Research Article Implementation of 2D Discrete Wavelet ...downloads.hindawi.com/journals/mpe/2014/532979.pdfResearch Article Implementation of 2D Discrete Wavelet Transform by Number
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Research ArticleImplementation of 2D Discrete Wavelet Transform by NumberTheoretic Transform and 2D Overlap-Save Method
Lina Yang12 Yuan Yan Tang1 and Qi Sun3
1 Department of Computer and Information Science University of Macau Avenida Padre Tomas Pereira Taipa 1356 Macau2Department of Mathematics and Computer Science Guangxi Normal University of Nationalities Chongzuo 532200 China3Department of Mathematics Sichuan University Chengdu Sichuan 610064 China
Correspondence should be addressed to Yuan Yan Tang yytangumacmo
Received 14 January 2014 Accepted 25 April 2014 Published 27 May 2014
Academic Editor Cristian Toma
Copyright copy 2014 Lina Yang et alThis is an open access article distributed under theCreativeCommonsAttributionLicensewhichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
To reduce the computation complexity of wavelet transform this paper presents a novel approach to be implemented It consists oftwo key techniques (1) fast number theoretic transform(FNTT) In the FNTT linear convolution is replaced by the circular one Itcan speed up the computation of 2D discrete wavelet transform (2) In two-dimensional overlap-save method directly calculatingthe FNTT to thewhole input sequencemaymeet two difficulties namely a bigmodulo obstructs the effective implementation of theFNTT and a long input sequence slows the computation of the FNTT down To fight with such deficiencies a new technique whichis referred to as 2D overlap-save method is developed Experiments have been conducted The fast number theoretic transformand 2D overlap-method have been used to implement the dyadic wavelet transform and applied to contour extraction in patternrecognition
1 Introduction
Wavelet transform plays an important role in image process-ing pattern recognition document analyses and so forth [1ndash12] The basic operation of it in fact is a linear convolution
In digital signal processing a signal is represented by adiscrete sequence Therefore the discrete wavelet transformcan be utilized to process it We have [13]
Explicitly either continuous wavelet transform or discretewavelet transform is essentially the operation of filtering
Directly calculating (1) may be time-consuming becausethe number of multiplication operations may come large inthis way Although Mallat algorithm [14] can implement thewavelet transform successfully in some special cases thatis in an orthogonal wavelet basis or in the multiresolutionanalysis it is not a solution of all kinds of wavelet transformsTherefore studying how to speed up the general discretewavelet transform is of great practical significance
Actually the filtering is a linear convolution in signalprocessing In the following without loss of generality wewill begin with the definition of Linear Convolution of two2D sequences
The equation for a 2D finite impulse response of a filtercan be represented in the similar way and given by
11991011989911198992
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
1198991minus1198961ge0 1198992minus1198962ge0
1198991= 0 1 119899
2= 0 1
(3)
Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2014 Article ID 532979 15 pageshttpdxdoiorg1011552014532979
2 Mathematical Problems in Engineering
where 11990911989911198992
and 11991011989911198992
(1198991
= 0 1 1198992
= 0 1 ) standfor the input and output of the filter respectively and ℎ
11989611198962
denote the filtering coefficients (1198961
= 0 1 1198972minus 1 119896
2=
0 1 1198982minus 1)
In many certain applications we are required to computeonly such an output sequence which is with a finite length in(3) Let 119909
11989911198992
be a 2D finite input of the filter with lengths 1198971
and 1198981 We compute the finite output as follows
11991011989911198992
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt1198971
0le1198992minus1198962lt1198981
1198991= 0 1 119897
1minus 1 119899
2= 0 1 119898
1minus 1
(4)
In practice the difference between the length of the inputsignal and that of the filter is often large If we calculateconvolution (4) directly a large number of multiplicationoperations will be executed In the meantime a series ofcalculations to treat zeros will be performed if fast Fouriertransform (FFT) is used to speed up the computation Bothof these twomethods will be time-consuming to compute thelinear convolution To overcome this problem in this paperwe will present a novel approach which is referred to ourprevious work as number theoretic transform (NTT) [15ndash17] Also we will prove that the computation of the linearconvolution can be replaced by that of the cyclic convolutionof two 2D sequences with lengths 119873
1= 1198971
+ 1198972
minus 1 and1198732= 1198981+ 1198982minus 1 respectively
For two 2D sequences 11990911989911198992
and ℎ11989911198992
of finite lengths1198731
and 1198732 their cyclic convolution can be written by
where ⟨119909⟩119873denotes the remainder 119903 of 119909 modulo 119873 that is
119909 = 119902119873 + 119903 119902 is any integer and 0 le 119903 lt 119873The number theoretic transform (NTT) provides an
effective way to calculate the cyclic convolution Howeverthere exist two difficulties if directly applying the NTT to thewhole input sequence namely
(i) a considerable big modulo M has to be imposed Itobstructs the effective implementation of the NTTdue to the limited length of the word which is usedto store data in computers It will thus become abottleneck in the computation of convolution
(iii) the advantage of the fast computation of the NTTcannot be reached when the difference between thelength of the input sequence and that of the filtersequence is large Thus it may not speed up thecalculation of convolution in this case
To avoid using a big modulo M we can use Chineseremainder theorem to reduce the length of the modulo Mand successfully apply the NTT to calculate the convolutionUnfortunately the number of multiplication operations willbe at least doubled and thus the computational complexitywill increase The second key technique will be worked outin this paper to overcome the above difficulties and speed upthe calculation of the convolution It is termed 2D overlap-save method which is an expansion of the 1D overlap-save method It can be used to implement the 2D wavelettransform when a big difference between the length of theinput sequence 119909
11989911198992
and that of the filter sequence ℎ11989911198992
occurs This method consists of three steps
(i) First the original 2D input sequence is equivalentlydivided into many small separated sections Thisdivision brings two evident improvements
(a) The size of each section is much smaller thanthat of the whole input sequence such that asmallmodulo M can be used and the NTT cantherefore be performed in a computer system
(b) The difference between the length of one sectionand that of the filter sequence becomes smallerwhich makes the effective application of theNTT
(ii) In the second step we calculate the cyclic convolutionof each section with the filter sequence by NTT
(iii) Finally the result of (4) is obtained by picking out datafrom each of the results calculated in the second stepand combining these data together
In comparison with direct method and fast Fouriertransform (FFT) it will be proved explicitly that the numberof multiplication operations in the 2D overlap-save methodwill be smaller than that in any of those twomethods inmanyparticular cases
In the next section the number theoretic transform(NTT) will be presented Details of the 2D overlap-methodwill be discussed in Section 3 A comparison of computationof the multiplication in three methods will be given inSection 4 A computational example will be presented aswell as a practical experiment in Section 5 which will verifythe efficiency of the proposed approach In the experimentthe fast number theoretic transform and 2D overlap-methodwere applied to implement the dyadic wavelet transformextracting the contours in recognition of Chinese handwrit-ing
2 Number Theoretic Transform (NTT)
Suppose that ℎ11989611198962
are the filter coefficients (1198961= 0 1 119897
2minus
1 1198962
= 0 1 1198982minus 1) 119909
11989911198992
and 11991011989911198992
(1198991
= 0 1 1198971minus
1 1198992
= 0 1 1198981
minus 1) are the input and output of thefilter respectively In the concrete applications of the discretewavelet transform (DWT) the lengths of the input and outputare very long (119897
1and 119898
1are very big) while lengths of the
filter coefficients are short (1198972and 119898
2are small) Therefore if
Mathematical Problems in Engineering 3
the directmethod is used to calculate their linear convolutionthe number of multiplication operations will be large Inthis section the 2D number theoretic transform (FNTT) willbe applied to calculate the 2D circular convolution of 2Dsequences of lengths119873
1= 1198971+1198972minus1 and119873
2= 1198981+1198982minus1 In a
2D number theoretic transform (NTT) if a modulo119872 is toobig the Chinese remainder theorem (CRT) will be utilized toreduce the length
21 2D Linear Convolution and 2D Circular ConvolutionGiven two 2-dimensional sequences 119909
11989911198992
(1198991= 0 1 119897
1minus
1 1198992= 0 1 119898
1minus 1) and ℎ
11989911198992
(1198991= 0 1 119897
2minus 1 119899
2=
0 1 1198982minus 1) their linear convolution is defined by
11991011989911198992
=
1198971minus1
sum
1198961=0
1198981minus1
sum
1198962=0
11990911989611198962
ℎ1198991minus11989611198992minus1198962
0le1198991minus1198961lt1198972
0le1198992minus1198962lt1198982
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt1198971
0le1198992minus1198962lt1198981
1198991= 0 1 119897
1+ 1198972minus 2 119899
2= 0 1 119898
1+ 1198982minus 2
(6)
Given two 2-dimensional sequences 11990911989911198992
are distinct primes then the congruence equation (8) is a 2-dimensional NTT mod119872 with the first dimension of length1198731and the second dimension of length119873
2 if and only if 1205721198731 equiv
1 (mod119872) 1205731198732 equiv 1 (mod119872) and 120572119895
equiv 1 (mod119901119896) (1 le
119895 le 1198731minus 1) 120573119894 equiv 1 (mod 119901
119896) (1 le 119894 le 119873
2minus 1) 119896 = 1 119904
Corollary 3 The congruence equation (8) is a 2-dimensionalNTT mod119872 if and only if
lcm [1198731 1198732] | gcd (119901
1minus 1 119901
119904minus 1) (13)
where lcm[1198731 1198732] denotes the least common multiple of two
integers1198731and119873
2and gcd(119886 119887) denotes the greatest common
divisor of two integers 119886 and 119887
Corollary 4 Let 119901 be an odd prime and 119872 = 119901 Thenthe congruence equation (8) is a 2-dimensional NTT mod119901 ifand only if 1205721198731 equiv 1 (mod119901) and 120572
119895equiv 1 (mod119901) (1 le 119895 le
1198731minus 1) 1205731198732 equiv 1 (mod119901) and 120573
119895equiv 1 (mod119901) (1 le 119895 le
1198732minus 1)
Clearly from congruence equations (8) (9) (11) and(12) we can use the 2-dimensional NTT to calculate a 2-dimensional circular convolution
23 Fast Number Theoretic Transform (FNTT) The idea ofFFT can be used to perform the NTT In this subsection atheoretic description will be presented briefly More detailswill be given in Section 5
Let
11988001198962
equiv
1198732minus1
sum
1198992=0
11990901198992
12057311989921198962 119880
1198731minus11198962
equiv
1198732minus1
sum
1198992=0
1199091198731minus11198992
12057311989921198962(mod119872)
1198962= 0 1 119873
2minus 1
(17)
From congruence equation (17) we deduce that
11988311989610
equiv
1198731minus1
sum
1198991=0
1198801198991012057211989911198961 119883
11989611198732minus1
equiv
1198731minus1
sum
1198991=0
11988011989911198732minus1
12057211989911198961(mod 119872)
1198961= 0 1 119873
1minus 1
(18)
Suppose that 1198731and 119873
2satisfy inequalities and we then
can use the idea of FFT to calculate the congruence equations(17) and (18)
For computing every
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod 119872)
(1198962= 0 1 119873
2minus 1)
(19)
in congruence equation (17) the numbers of all multipli-cation necessary are (119873
2log1198732)2 by using FFT algorithm
Hence if using FFT algorithm to calculate the congruence
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod119872)
(1198962= 0 1 119873
2minus 1)
(20)
in congruence equation (17) the numbers of all multiplica-tion necessary are ((119873
11198732)2) log119873
2
Similarly if using FFT algorithm to calculate all congru-ence equations in (18) then the numbers of all multiplicationnecessary are ((119873
21198731)2) log119873
1
Therefore if using the fast number theoretic transform(FNTT) to calculate the congruence equation (8) the numberof all multiplication necessary is
11987311198732
2log1198732+
11987321198731
2log1198731=
11987311198732
2log (119873
11198732) (21)
If119872 is very large wemay reduce the length of aword by usingChinese remainder theorem and it can be abbreviated toCRT
Therefore we have the following proposition
Proposition 5 Suppose that congruence equation (8) is a 2-dimensional NTT mod119872 where 119872 = 119901
1198971
1sdot sdot sdot 119901119897119904
119904 1199011 119901
119904
are distinct primes then we have a total of s 2-dimensionalNTT mod 119901
119897119894
119894 and they are described as follows
11988311989611198962
equiv
1198731minus1
sum
1198991=0
1198732minus1
sum
1198992=0
11990911989911198992
12057211989911198961
11989412057311989921198962
119894 (mod 119901
119897119894
119894)
1198961= 0 1 119873
1minus 1 119896
2= 0 1 119873
2minus 1
(22)
where 120572119894= ⟨120572⟩
119901119897119894
119894
120573119894= ⟨120573⟩
119901119897119894
119894
119894 = 1 2 119904If
11991011989911198992
=
1198731minus1
sum
1198961=0
1198732minus1
sum
1198962=0
11990911989611198962
ℎ⟨1198991minus1198961⟩1198731
⟨1198992minus1198962⟩1198732
equiv 119910(119894)
11989911198992
(mod 119901119897119894
119894)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(23)
then
11991011989911198992
equiv
119904
sum
119894=1
1198721015840
119894119872119894119910(119894)
11989911198992
(mod119872)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(24)
where 119872 = 119901119897119894
1198941198721198941198721015840
119894119872119894equiv 1 (mod 119901
119897119894
119894) 119894 = 1 119904
Mathematical Problems in Engineering 5
Proof Because the congruence equation (8) is a 2-dimen-sional NTT mod119872 hence
1205721198731 equiv 1 (mod119872) 120573
1198732 equiv 1 (mod 119872)
1205721198731 equiv 1 (mod 119901
119894) 120572
119895equiv 1 (mod 119901
119894) 1 le 119895 lt 119873
1
1205731198732 equiv 1 (mod 119901
119894) 120573
ℎequiv (mod 119901
119894) 1 le ℎ lt 119873
2
119894 = 1 2 119904
(25)
so
1205721198731 equiv 1 (mod 119901
119897119894
119894) 120573
1198732 equiv 1 (mod 119901
119897119894
119894)
1205721198731
119894equiv 1205721198731 equiv 1 (mod 119901
119894) 120572
119895
119894equiv 1 (mod 119901
119894)
1 le 119895 lt 1198731
1205731198732
119894equiv 1205731198732 equiv 1 (mod 119901
119894) 120573
ℎ
119894equiv 1 (mod 119901
119894)
1 le ℎ lt 1198732
119894 = 1 119904
(26)
Hence the congruence equation (22) has a total of 119904 2-dimensional NTTs mod 119901
119897119894
119894 119894 = 1 119904 From CRT and the
congruence equation (23) we deduce that the congruenceequation (24) holds
3 Two-Dimensional Overlap-Save Method
Select two positive integers 1198731015840 and 1198721015840 so that 1198731015840 + 119897
2minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 either 119873 or 119872 is given
by an integer power of 2 that is 119873 = 21198891 and 119872 = 2
1198892 where
1198891and 119889
2are two positive integers
Assuming that the input sequence 11990911989911198992
(1198991
=
0 1 1198971
minus 1 1198992
= 0 1 1198981
minus 1) is divided intomany small sections which are termed submatrices
The consecutive matrix is overlapped by the previous oneAs a simple example four blocks chosen from the dividedinput matrices are overlapped and are graphically shown inFigure 1 Parameters V
1and V
2can be considered as shift
parameters because changing either V1or V2will select dif-
ferent submatrix For the first submatrix shown in Figure 1(a)we select V
1= 0 and V
2= 0 Second submatrix (V
1= 0 and
V2= 1) is displayed in Figure 1(b)Third one with parameters
of V1
= 1 and V2
= 0 is given in Figure 1(c) Fourth matrixshown in Figure 1(d) has parameters V
1= 1 and V
2= 1 The
procedure of the overlapping is also graphically illustrated inFigure 1
For the operation of circular convolution it is necessaryto have two sequences with the same length Taking (7) for
example the sequences 11990911989911198992
and ℎ11989911198992
are of the same lengthof1198731times1198732 Since the input sequence has already divided into
submatrices of119873times119872 by (27) the filter sequence has to be oflength119873times119872Therefore119873minus119897
2augmenting zeros are required
to be added to the row and119872minus1198982zeros to the column of the
filter sequence ℎ11989911198992
As a result the sequence ℎ11989911198992
becomes
ℎ1015840
11989911198992
=ℎ11989911198992
if 1198991= 0 1 119897
2minus 1 119899
2= 0 1 119898
2minus 1
0 if 1198972le 1198991lt 119873 or 119898
2le 1198992lt 119872
(28)
where 1198991= 0 1 119873 minus 1 and 119899
2= 0 1 119872 minus 1
The following form shows how a sequence ℎ101584011989911198992
is yieldedby adding augmenting zeros to the filter sequence ℎ
11989911198992
ℎ 997904rArr ℎ1015840 =
ℎ
0 (29)
For each of the divided submatrices we can give V1and V
2
and compute the circular convolution of two 2D sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
The result of (31) is that of circular convolution As we willsee below there exists an important connection between thecircular convolution and the linear one In fact since 119899
1=
0 1 1198972minus 1 and 119899
2= 0 1 119898
2minus 1 we can choose 119905
1=
1198972minus1 119873minus1 and 119905
2= 1198982minus1 119872minus1 in (31) this makes
119873 gt 1199051minus 1198991
gt 0 and 119872 gt 1199052minus 1198992
gt 0 in (31) Thus theremainder can exactly be equal to 119905
In order to clearly understand the procedure of obtainingthe components of linear convolution from the circularconvolution equation (31) a graphic description of it can befound in Figure 2 We show a dashed rectangle with size of119873 by 119872 in Figure 2(a) to represent the circular convolutionwhich is described in (31) and a darkened rectangle with sizeof1198731015840 by119872
1015840 to indicate the linear convolution stated in (32)
6 Mathematical Problems in Engineering
2N
m
n
2M
M
N1 = 0 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N1 = 1 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
m
n
2N
2M
M
N
1 = 0 2 = 1
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(c)
m
n
2N
2M
M
N
1 = 1 2 = 1l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(d)
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
= 0 1 ) standfor the input and output of the filter respectively and ℎ
11989611198962
denote the filtering coefficients (1198961
= 0 1 1198972minus 1 119896
2=
0 1 1198982minus 1)
In many certain applications we are required to computeonly such an output sequence which is with a finite length in(3) Let 119909
11989911198992
be a 2D finite input of the filter with lengths 1198971
and 1198981 We compute the finite output as follows
11991011989911198992
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt1198971
0le1198992minus1198962lt1198981
1198991= 0 1 119897
1minus 1 119899
2= 0 1 119898
1minus 1
(4)
In practice the difference between the length of the inputsignal and that of the filter is often large If we calculateconvolution (4) directly a large number of multiplicationoperations will be executed In the meantime a series ofcalculations to treat zeros will be performed if fast Fouriertransform (FFT) is used to speed up the computation Bothof these twomethods will be time-consuming to compute thelinear convolution To overcome this problem in this paperwe will present a novel approach which is referred to ourprevious work as number theoretic transform (NTT) [15ndash17] Also we will prove that the computation of the linearconvolution can be replaced by that of the cyclic convolutionof two 2D sequences with lengths 119873
1= 1198971
+ 1198972
minus 1 and1198732= 1198981+ 1198982minus 1 respectively
For two 2D sequences 11990911989911198992
and ℎ11989911198992
of finite lengths1198731
and 1198732 their cyclic convolution can be written by
where ⟨119909⟩119873denotes the remainder 119903 of 119909 modulo 119873 that is
119909 = 119902119873 + 119903 119902 is any integer and 0 le 119903 lt 119873The number theoretic transform (NTT) provides an
effective way to calculate the cyclic convolution Howeverthere exist two difficulties if directly applying the NTT to thewhole input sequence namely
(i) a considerable big modulo M has to be imposed Itobstructs the effective implementation of the NTTdue to the limited length of the word which is usedto store data in computers It will thus become abottleneck in the computation of convolution
(iii) the advantage of the fast computation of the NTTcannot be reached when the difference between thelength of the input sequence and that of the filtersequence is large Thus it may not speed up thecalculation of convolution in this case
To avoid using a big modulo M we can use Chineseremainder theorem to reduce the length of the modulo Mand successfully apply the NTT to calculate the convolutionUnfortunately the number of multiplication operations willbe at least doubled and thus the computational complexitywill increase The second key technique will be worked outin this paper to overcome the above difficulties and speed upthe calculation of the convolution It is termed 2D overlap-save method which is an expansion of the 1D overlap-save method It can be used to implement the 2D wavelettransform when a big difference between the length of theinput sequence 119909
11989911198992
and that of the filter sequence ℎ11989911198992
occurs This method consists of three steps
(i) First the original 2D input sequence is equivalentlydivided into many small separated sections Thisdivision brings two evident improvements
(a) The size of each section is much smaller thanthat of the whole input sequence such that asmallmodulo M can be used and the NTT cantherefore be performed in a computer system
(b) The difference between the length of one sectionand that of the filter sequence becomes smallerwhich makes the effective application of theNTT
(ii) In the second step we calculate the cyclic convolutionof each section with the filter sequence by NTT
(iii) Finally the result of (4) is obtained by picking out datafrom each of the results calculated in the second stepand combining these data together
In comparison with direct method and fast Fouriertransform (FFT) it will be proved explicitly that the numberof multiplication operations in the 2D overlap-save methodwill be smaller than that in any of those twomethods inmanyparticular cases
In the next section the number theoretic transform(NTT) will be presented Details of the 2D overlap-methodwill be discussed in Section 3 A comparison of computationof the multiplication in three methods will be given inSection 4 A computational example will be presented aswell as a practical experiment in Section 5 which will verifythe efficiency of the proposed approach In the experimentthe fast number theoretic transform and 2D overlap-methodwere applied to implement the dyadic wavelet transformextracting the contours in recognition of Chinese handwrit-ing
2 Number Theoretic Transform (NTT)
Suppose that ℎ11989611198962
are the filter coefficients (1198961= 0 1 119897
2minus
1 1198962
= 0 1 1198982minus 1) 119909
11989911198992
and 11991011989911198992
(1198991
= 0 1 1198971minus
1 1198992
= 0 1 1198981
minus 1) are the input and output of thefilter respectively In the concrete applications of the discretewavelet transform (DWT) the lengths of the input and outputare very long (119897
1and 119898
1are very big) while lengths of the
filter coefficients are short (1198972and 119898
2are small) Therefore if
Mathematical Problems in Engineering 3
the directmethod is used to calculate their linear convolutionthe number of multiplication operations will be large Inthis section the 2D number theoretic transform (FNTT) willbe applied to calculate the 2D circular convolution of 2Dsequences of lengths119873
1= 1198971+1198972minus1 and119873
2= 1198981+1198982minus1 In a
2D number theoretic transform (NTT) if a modulo119872 is toobig the Chinese remainder theorem (CRT) will be utilized toreduce the length
21 2D Linear Convolution and 2D Circular ConvolutionGiven two 2-dimensional sequences 119909
11989911198992
(1198991= 0 1 119897
1minus
1 1198992= 0 1 119898
1minus 1) and ℎ
11989911198992
(1198991= 0 1 119897
2minus 1 119899
2=
0 1 1198982minus 1) their linear convolution is defined by
11991011989911198992
=
1198971minus1
sum
1198961=0
1198981minus1
sum
1198962=0
11990911989611198962
ℎ1198991minus11989611198992minus1198962
0le1198991minus1198961lt1198972
0le1198992minus1198962lt1198982
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt1198971
0le1198992minus1198962lt1198981
1198991= 0 1 119897
1+ 1198972minus 2 119899
2= 0 1 119898
1+ 1198982minus 2
(6)
Given two 2-dimensional sequences 11990911989911198992
are distinct primes then the congruence equation (8) is a 2-dimensional NTT mod119872 with the first dimension of length1198731and the second dimension of length119873
2 if and only if 1205721198731 equiv
1 (mod119872) 1205731198732 equiv 1 (mod119872) and 120572119895
equiv 1 (mod119901119896) (1 le
119895 le 1198731minus 1) 120573119894 equiv 1 (mod 119901
119896) (1 le 119894 le 119873
2minus 1) 119896 = 1 119904
Corollary 3 The congruence equation (8) is a 2-dimensionalNTT mod119872 if and only if
lcm [1198731 1198732] | gcd (119901
1minus 1 119901
119904minus 1) (13)
where lcm[1198731 1198732] denotes the least common multiple of two
integers1198731and119873
2and gcd(119886 119887) denotes the greatest common
divisor of two integers 119886 and 119887
Corollary 4 Let 119901 be an odd prime and 119872 = 119901 Thenthe congruence equation (8) is a 2-dimensional NTT mod119901 ifand only if 1205721198731 equiv 1 (mod119901) and 120572
119895equiv 1 (mod119901) (1 le 119895 le
1198731minus 1) 1205731198732 equiv 1 (mod119901) and 120573
119895equiv 1 (mod119901) (1 le 119895 le
1198732minus 1)
Clearly from congruence equations (8) (9) (11) and(12) we can use the 2-dimensional NTT to calculate a 2-dimensional circular convolution
23 Fast Number Theoretic Transform (FNTT) The idea ofFFT can be used to perform the NTT In this subsection atheoretic description will be presented briefly More detailswill be given in Section 5
Let
11988001198962
equiv
1198732minus1
sum
1198992=0
11990901198992
12057311989921198962 119880
1198731minus11198962
equiv
1198732minus1
sum
1198992=0
1199091198731minus11198992
12057311989921198962(mod119872)
1198962= 0 1 119873
2minus 1
(17)
From congruence equation (17) we deduce that
11988311989610
equiv
1198731minus1
sum
1198991=0
1198801198991012057211989911198961 119883
11989611198732minus1
equiv
1198731minus1
sum
1198991=0
11988011989911198732minus1
12057211989911198961(mod 119872)
1198961= 0 1 119873
1minus 1
(18)
Suppose that 1198731and 119873
2satisfy inequalities and we then
can use the idea of FFT to calculate the congruence equations(17) and (18)
For computing every
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod 119872)
(1198962= 0 1 119873
2minus 1)
(19)
in congruence equation (17) the numbers of all multipli-cation necessary are (119873
2log1198732)2 by using FFT algorithm
Hence if using FFT algorithm to calculate the congruence
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod119872)
(1198962= 0 1 119873
2minus 1)
(20)
in congruence equation (17) the numbers of all multiplica-tion necessary are ((119873
11198732)2) log119873
2
Similarly if using FFT algorithm to calculate all congru-ence equations in (18) then the numbers of all multiplicationnecessary are ((119873
21198731)2) log119873
1
Therefore if using the fast number theoretic transform(FNTT) to calculate the congruence equation (8) the numberof all multiplication necessary is
11987311198732
2log1198732+
11987321198731
2log1198731=
11987311198732
2log (119873
11198732) (21)
If119872 is very large wemay reduce the length of aword by usingChinese remainder theorem and it can be abbreviated toCRT
Therefore we have the following proposition
Proposition 5 Suppose that congruence equation (8) is a 2-dimensional NTT mod119872 where 119872 = 119901
1198971
1sdot sdot sdot 119901119897119904
119904 1199011 119901
119904
are distinct primes then we have a total of s 2-dimensionalNTT mod 119901
119897119894
119894 and they are described as follows
11988311989611198962
equiv
1198731minus1
sum
1198991=0
1198732minus1
sum
1198992=0
11990911989911198992
12057211989911198961
11989412057311989921198962
119894 (mod 119901
119897119894
119894)
1198961= 0 1 119873
1minus 1 119896
2= 0 1 119873
2minus 1
(22)
where 120572119894= ⟨120572⟩
119901119897119894
119894
120573119894= ⟨120573⟩
119901119897119894
119894
119894 = 1 2 119904If
11991011989911198992
=
1198731minus1
sum
1198961=0
1198732minus1
sum
1198962=0
11990911989611198962
ℎ⟨1198991minus1198961⟩1198731
⟨1198992minus1198962⟩1198732
equiv 119910(119894)
11989911198992
(mod 119901119897119894
119894)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(23)
then
11991011989911198992
equiv
119904
sum
119894=1
1198721015840
119894119872119894119910(119894)
11989911198992
(mod119872)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(24)
where 119872 = 119901119897119894
1198941198721198941198721015840
119894119872119894equiv 1 (mod 119901
119897119894
119894) 119894 = 1 119904
Mathematical Problems in Engineering 5
Proof Because the congruence equation (8) is a 2-dimen-sional NTT mod119872 hence
1205721198731 equiv 1 (mod119872) 120573
1198732 equiv 1 (mod 119872)
1205721198731 equiv 1 (mod 119901
119894) 120572
119895equiv 1 (mod 119901
119894) 1 le 119895 lt 119873
1
1205731198732 equiv 1 (mod 119901
119894) 120573
ℎequiv (mod 119901
119894) 1 le ℎ lt 119873
2
119894 = 1 2 119904
(25)
so
1205721198731 equiv 1 (mod 119901
119897119894
119894) 120573
1198732 equiv 1 (mod 119901
119897119894
119894)
1205721198731
119894equiv 1205721198731 equiv 1 (mod 119901
119894) 120572
119895
119894equiv 1 (mod 119901
119894)
1 le 119895 lt 1198731
1205731198732
119894equiv 1205731198732 equiv 1 (mod 119901
119894) 120573
ℎ
119894equiv 1 (mod 119901
119894)
1 le ℎ lt 1198732
119894 = 1 119904
(26)
Hence the congruence equation (22) has a total of 119904 2-dimensional NTTs mod 119901
119897119894
119894 119894 = 1 119904 From CRT and the
congruence equation (23) we deduce that the congruenceequation (24) holds
3 Two-Dimensional Overlap-Save Method
Select two positive integers 1198731015840 and 1198721015840 so that 1198731015840 + 119897
2minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 either 119873 or 119872 is given
by an integer power of 2 that is 119873 = 21198891 and 119872 = 2
1198892 where
1198891and 119889
2are two positive integers
Assuming that the input sequence 11990911989911198992
(1198991
=
0 1 1198971
minus 1 1198992
= 0 1 1198981
minus 1) is divided intomany small sections which are termed submatrices
The consecutive matrix is overlapped by the previous oneAs a simple example four blocks chosen from the dividedinput matrices are overlapped and are graphically shown inFigure 1 Parameters V
1and V
2can be considered as shift
parameters because changing either V1or V2will select dif-
ferent submatrix For the first submatrix shown in Figure 1(a)we select V
1= 0 and V
2= 0 Second submatrix (V
1= 0 and
V2= 1) is displayed in Figure 1(b)Third one with parameters
of V1
= 1 and V2
= 0 is given in Figure 1(c) Fourth matrixshown in Figure 1(d) has parameters V
1= 1 and V
2= 1 The
procedure of the overlapping is also graphically illustrated inFigure 1
For the operation of circular convolution it is necessaryto have two sequences with the same length Taking (7) for
example the sequences 11990911989911198992
and ℎ11989911198992
are of the same lengthof1198731times1198732 Since the input sequence has already divided into
submatrices of119873times119872 by (27) the filter sequence has to be oflength119873times119872Therefore119873minus119897
2augmenting zeros are required
to be added to the row and119872minus1198982zeros to the column of the
filter sequence ℎ11989911198992
As a result the sequence ℎ11989911198992
becomes
ℎ1015840
11989911198992
=ℎ11989911198992
if 1198991= 0 1 119897
2minus 1 119899
2= 0 1 119898
2minus 1
0 if 1198972le 1198991lt 119873 or 119898
2le 1198992lt 119872
(28)
where 1198991= 0 1 119873 minus 1 and 119899
2= 0 1 119872 minus 1
The following form shows how a sequence ℎ101584011989911198992
is yieldedby adding augmenting zeros to the filter sequence ℎ
11989911198992
ℎ 997904rArr ℎ1015840 =
ℎ
0 (29)
For each of the divided submatrices we can give V1and V
2
and compute the circular convolution of two 2D sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
The result of (31) is that of circular convolution As we willsee below there exists an important connection between thecircular convolution and the linear one In fact since 119899
1=
0 1 1198972minus 1 and 119899
2= 0 1 119898
2minus 1 we can choose 119905
1=
1198972minus1 119873minus1 and 119905
2= 1198982minus1 119872minus1 in (31) this makes
119873 gt 1199051minus 1198991
gt 0 and 119872 gt 1199052minus 1198992
gt 0 in (31) Thus theremainder can exactly be equal to 119905
In order to clearly understand the procedure of obtainingthe components of linear convolution from the circularconvolution equation (31) a graphic description of it can befound in Figure 2 We show a dashed rectangle with size of119873 by 119872 in Figure 2(a) to represent the circular convolutionwhich is described in (31) and a darkened rectangle with sizeof1198731015840 by119872
1015840 to indicate the linear convolution stated in (32)
6 Mathematical Problems in Engineering
2N
m
n
2M
M
N1 = 0 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N1 = 1 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
m
n
2N
2M
M
N
1 = 0 2 = 1
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(c)
m
n
2N
2M
M
N
1 = 1 2 = 1l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(d)
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
the directmethod is used to calculate their linear convolutionthe number of multiplication operations will be large Inthis section the 2D number theoretic transform (FNTT) willbe applied to calculate the 2D circular convolution of 2Dsequences of lengths119873
1= 1198971+1198972minus1 and119873
2= 1198981+1198982minus1 In a
2D number theoretic transform (NTT) if a modulo119872 is toobig the Chinese remainder theorem (CRT) will be utilized toreduce the length
21 2D Linear Convolution and 2D Circular ConvolutionGiven two 2-dimensional sequences 119909
11989911198992
(1198991= 0 1 119897
1minus
1 1198992= 0 1 119898
1minus 1) and ℎ
11989911198992
(1198991= 0 1 119897
2minus 1 119899
2=
0 1 1198982minus 1) their linear convolution is defined by
11991011989911198992
=
1198971minus1
sum
1198961=0
1198981minus1
sum
1198962=0
11990911989611198962
ℎ1198991minus11989611198992minus1198962
0le1198991minus1198961lt1198972
0le1198992minus1198962lt1198982
=
1198972minus1
sum
1198961=0
1198982minus1
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt1198971
0le1198992minus1198962lt1198981
1198991= 0 1 119897
1+ 1198972minus 2 119899
2= 0 1 119898
1+ 1198982minus 2
(6)
Given two 2-dimensional sequences 11990911989911198992
are distinct primes then the congruence equation (8) is a 2-dimensional NTT mod119872 with the first dimension of length1198731and the second dimension of length119873
2 if and only if 1205721198731 equiv
1 (mod119872) 1205731198732 equiv 1 (mod119872) and 120572119895
equiv 1 (mod119901119896) (1 le
119895 le 1198731minus 1) 120573119894 equiv 1 (mod 119901
119896) (1 le 119894 le 119873
2minus 1) 119896 = 1 119904
Corollary 3 The congruence equation (8) is a 2-dimensionalNTT mod119872 if and only if
lcm [1198731 1198732] | gcd (119901
1minus 1 119901
119904minus 1) (13)
where lcm[1198731 1198732] denotes the least common multiple of two
integers1198731and119873
2and gcd(119886 119887) denotes the greatest common
divisor of two integers 119886 and 119887
Corollary 4 Let 119901 be an odd prime and 119872 = 119901 Thenthe congruence equation (8) is a 2-dimensional NTT mod119901 ifand only if 1205721198731 equiv 1 (mod119901) and 120572
119895equiv 1 (mod119901) (1 le 119895 le
1198731minus 1) 1205731198732 equiv 1 (mod119901) and 120573
119895equiv 1 (mod119901) (1 le 119895 le
1198732minus 1)
Clearly from congruence equations (8) (9) (11) and(12) we can use the 2-dimensional NTT to calculate a 2-dimensional circular convolution
23 Fast Number Theoretic Transform (FNTT) The idea ofFFT can be used to perform the NTT In this subsection atheoretic description will be presented briefly More detailswill be given in Section 5
Let
11988001198962
equiv
1198732minus1
sum
1198992=0
11990901198992
12057311989921198962 119880
1198731minus11198962
equiv
1198732minus1
sum
1198992=0
1199091198731minus11198992
12057311989921198962(mod119872)
1198962= 0 1 119873
2minus 1
(17)
From congruence equation (17) we deduce that
11988311989610
equiv
1198731minus1
sum
1198991=0
1198801198991012057211989911198961 119883
11989611198732minus1
equiv
1198731minus1
sum
1198991=0
11988011989911198732minus1
12057211989911198961(mod 119872)
1198961= 0 1 119873
1minus 1
(18)
Suppose that 1198731and 119873
2satisfy inequalities and we then
can use the idea of FFT to calculate the congruence equations(17) and (18)
For computing every
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod 119872)
(1198962= 0 1 119873
2minus 1)
(19)
in congruence equation (17) the numbers of all multipli-cation necessary are (119873
2log1198732)2 by using FFT algorithm
Hence if using FFT algorithm to calculate the congruence
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod119872)
(1198962= 0 1 119873
2minus 1)
(20)
in congruence equation (17) the numbers of all multiplica-tion necessary are ((119873
11198732)2) log119873
2
Similarly if using FFT algorithm to calculate all congru-ence equations in (18) then the numbers of all multiplicationnecessary are ((119873
21198731)2) log119873
1
Therefore if using the fast number theoretic transform(FNTT) to calculate the congruence equation (8) the numberof all multiplication necessary is
11987311198732
2log1198732+
11987321198731
2log1198731=
11987311198732
2log (119873
11198732) (21)
If119872 is very large wemay reduce the length of aword by usingChinese remainder theorem and it can be abbreviated toCRT
Therefore we have the following proposition
Proposition 5 Suppose that congruence equation (8) is a 2-dimensional NTT mod119872 where 119872 = 119901
1198971
1sdot sdot sdot 119901119897119904
119904 1199011 119901
119904
are distinct primes then we have a total of s 2-dimensionalNTT mod 119901
119897119894
119894 and they are described as follows
11988311989611198962
equiv
1198731minus1
sum
1198991=0
1198732minus1
sum
1198992=0
11990911989911198992
12057211989911198961
11989412057311989921198962
119894 (mod 119901
119897119894
119894)
1198961= 0 1 119873
1minus 1 119896
2= 0 1 119873
2minus 1
(22)
where 120572119894= ⟨120572⟩
119901119897119894
119894
120573119894= ⟨120573⟩
119901119897119894
119894
119894 = 1 2 119904If
11991011989911198992
=
1198731minus1
sum
1198961=0
1198732minus1
sum
1198962=0
11990911989611198962
ℎ⟨1198991minus1198961⟩1198731
⟨1198992minus1198962⟩1198732
equiv 119910(119894)
11989911198992
(mod 119901119897119894
119894)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(23)
then
11991011989911198992
equiv
119904
sum
119894=1
1198721015840
119894119872119894119910(119894)
11989911198992
(mod119872)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(24)
where 119872 = 119901119897119894
1198941198721198941198721015840
119894119872119894equiv 1 (mod 119901
119897119894
119894) 119894 = 1 119904
Mathematical Problems in Engineering 5
Proof Because the congruence equation (8) is a 2-dimen-sional NTT mod119872 hence
1205721198731 equiv 1 (mod119872) 120573
1198732 equiv 1 (mod 119872)
1205721198731 equiv 1 (mod 119901
119894) 120572
119895equiv 1 (mod 119901
119894) 1 le 119895 lt 119873
1
1205731198732 equiv 1 (mod 119901
119894) 120573
ℎequiv (mod 119901
119894) 1 le ℎ lt 119873
2
119894 = 1 2 119904
(25)
so
1205721198731 equiv 1 (mod 119901
119897119894
119894) 120573
1198732 equiv 1 (mod 119901
119897119894
119894)
1205721198731
119894equiv 1205721198731 equiv 1 (mod 119901
119894) 120572
119895
119894equiv 1 (mod 119901
119894)
1 le 119895 lt 1198731
1205731198732
119894equiv 1205731198732 equiv 1 (mod 119901
119894) 120573
ℎ
119894equiv 1 (mod 119901
119894)
1 le ℎ lt 1198732
119894 = 1 119904
(26)
Hence the congruence equation (22) has a total of 119904 2-dimensional NTTs mod 119901
119897119894
119894 119894 = 1 119904 From CRT and the
congruence equation (23) we deduce that the congruenceequation (24) holds
3 Two-Dimensional Overlap-Save Method
Select two positive integers 1198731015840 and 1198721015840 so that 1198731015840 + 119897
2minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 either 119873 or 119872 is given
by an integer power of 2 that is 119873 = 21198891 and 119872 = 2
1198892 where
1198891and 119889
2are two positive integers
Assuming that the input sequence 11990911989911198992
(1198991
=
0 1 1198971
minus 1 1198992
= 0 1 1198981
minus 1) is divided intomany small sections which are termed submatrices
The consecutive matrix is overlapped by the previous oneAs a simple example four blocks chosen from the dividedinput matrices are overlapped and are graphically shown inFigure 1 Parameters V
1and V
2can be considered as shift
parameters because changing either V1or V2will select dif-
ferent submatrix For the first submatrix shown in Figure 1(a)we select V
1= 0 and V
2= 0 Second submatrix (V
1= 0 and
V2= 1) is displayed in Figure 1(b)Third one with parameters
of V1
= 1 and V2
= 0 is given in Figure 1(c) Fourth matrixshown in Figure 1(d) has parameters V
1= 1 and V
2= 1 The
procedure of the overlapping is also graphically illustrated inFigure 1
For the operation of circular convolution it is necessaryto have two sequences with the same length Taking (7) for
example the sequences 11990911989911198992
and ℎ11989911198992
are of the same lengthof1198731times1198732 Since the input sequence has already divided into
submatrices of119873times119872 by (27) the filter sequence has to be oflength119873times119872Therefore119873minus119897
2augmenting zeros are required
to be added to the row and119872minus1198982zeros to the column of the
filter sequence ℎ11989911198992
As a result the sequence ℎ11989911198992
becomes
ℎ1015840
11989911198992
=ℎ11989911198992
if 1198991= 0 1 119897
2minus 1 119899
2= 0 1 119898
2minus 1
0 if 1198972le 1198991lt 119873 or 119898
2le 1198992lt 119872
(28)
where 1198991= 0 1 119873 minus 1 and 119899
2= 0 1 119872 minus 1
The following form shows how a sequence ℎ101584011989911198992
is yieldedby adding augmenting zeros to the filter sequence ℎ
11989911198992
ℎ 997904rArr ℎ1015840 =
ℎ
0 (29)
For each of the divided submatrices we can give V1and V
2
and compute the circular convolution of two 2D sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
The result of (31) is that of circular convolution As we willsee below there exists an important connection between thecircular convolution and the linear one In fact since 119899
1=
0 1 1198972minus 1 and 119899
2= 0 1 119898
2minus 1 we can choose 119905
1=
1198972minus1 119873minus1 and 119905
2= 1198982minus1 119872minus1 in (31) this makes
119873 gt 1199051minus 1198991
gt 0 and 119872 gt 1199052minus 1198992
gt 0 in (31) Thus theremainder can exactly be equal to 119905
In order to clearly understand the procedure of obtainingthe components of linear convolution from the circularconvolution equation (31) a graphic description of it can befound in Figure 2 We show a dashed rectangle with size of119873 by 119872 in Figure 2(a) to represent the circular convolutionwhich is described in (31) and a darkened rectangle with sizeof1198731015840 by119872
1015840 to indicate the linear convolution stated in (32)
6 Mathematical Problems in Engineering
2N
m
n
2M
M
N1 = 0 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N1 = 1 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
m
n
2N
2M
M
N
1 = 0 2 = 1
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(c)
m
n
2N
2M
M
N
1 = 1 2 = 1l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(d)
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
23 Fast Number Theoretic Transform (FNTT) The idea ofFFT can be used to perform the NTT In this subsection atheoretic description will be presented briefly More detailswill be given in Section 5
Let
11988001198962
equiv
1198732minus1
sum
1198992=0
11990901198992
12057311989921198962 119880
1198731minus11198962
equiv
1198732minus1
sum
1198992=0
1199091198731minus11198992
12057311989921198962(mod119872)
1198962= 0 1 119873
2minus 1
(17)
From congruence equation (17) we deduce that
11988311989610
equiv
1198731minus1
sum
1198991=0
1198801198991012057211989911198961 119883
11989611198732minus1
equiv
1198731minus1
sum
1198991=0
11988011989911198732minus1
12057211989911198961(mod 119872)
1198961= 0 1 119873
1minus 1
(18)
Suppose that 1198731and 119873
2satisfy inequalities and we then
can use the idea of FFT to calculate the congruence equations(17) and (18)
For computing every
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod 119872)
(1198962= 0 1 119873
2minus 1)
(19)
in congruence equation (17) the numbers of all multipli-cation necessary are (119873
2log1198732)2 by using FFT algorithm
Hence if using FFT algorithm to calculate the congruence
11988011989911198962
equiv
1198732minus1
sum
1198992=0
11990911989911198992
12057311989921198962(mod119872)
(1198962= 0 1 119873
2minus 1)
(20)
in congruence equation (17) the numbers of all multiplica-tion necessary are ((119873
11198732)2) log119873
2
Similarly if using FFT algorithm to calculate all congru-ence equations in (18) then the numbers of all multiplicationnecessary are ((119873
21198731)2) log119873
1
Therefore if using the fast number theoretic transform(FNTT) to calculate the congruence equation (8) the numberof all multiplication necessary is
11987311198732
2log1198732+
11987321198731
2log1198731=
11987311198732
2log (119873
11198732) (21)
If119872 is very large wemay reduce the length of aword by usingChinese remainder theorem and it can be abbreviated toCRT
Therefore we have the following proposition
Proposition 5 Suppose that congruence equation (8) is a 2-dimensional NTT mod119872 where 119872 = 119901
1198971
1sdot sdot sdot 119901119897119904
119904 1199011 119901
119904
are distinct primes then we have a total of s 2-dimensionalNTT mod 119901
119897119894
119894 and they are described as follows
11988311989611198962
equiv
1198731minus1
sum
1198991=0
1198732minus1
sum
1198992=0
11990911989911198992
12057211989911198961
11989412057311989921198962
119894 (mod 119901
119897119894
119894)
1198961= 0 1 119873
1minus 1 119896
2= 0 1 119873
2minus 1
(22)
where 120572119894= ⟨120572⟩
119901119897119894
119894
120573119894= ⟨120573⟩
119901119897119894
119894
119894 = 1 2 119904If
11991011989911198992
=
1198731minus1
sum
1198961=0
1198732minus1
sum
1198962=0
11990911989611198962
ℎ⟨1198991minus1198961⟩1198731
⟨1198992minus1198962⟩1198732
equiv 119910(119894)
11989911198992
(mod 119901119897119894
119894)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(23)
then
11991011989911198992
equiv
119904
sum
119894=1
1198721015840
119894119872119894119910(119894)
11989911198992
(mod119872)
(1198991= 0 1 119873
1minus 1 119899
2= 0 1 119873
2minus 1)
(24)
where 119872 = 119901119897119894
1198941198721198941198721015840
119894119872119894equiv 1 (mod 119901
119897119894
119894) 119894 = 1 119904
Mathematical Problems in Engineering 5
Proof Because the congruence equation (8) is a 2-dimen-sional NTT mod119872 hence
1205721198731 equiv 1 (mod119872) 120573
1198732 equiv 1 (mod 119872)
1205721198731 equiv 1 (mod 119901
119894) 120572
119895equiv 1 (mod 119901
119894) 1 le 119895 lt 119873
1
1205731198732 equiv 1 (mod 119901
119894) 120573
ℎequiv (mod 119901
119894) 1 le ℎ lt 119873
2
119894 = 1 2 119904
(25)
so
1205721198731 equiv 1 (mod 119901
119897119894
119894) 120573
1198732 equiv 1 (mod 119901
119897119894
119894)
1205721198731
119894equiv 1205721198731 equiv 1 (mod 119901
119894) 120572
119895
119894equiv 1 (mod 119901
119894)
1 le 119895 lt 1198731
1205731198732
119894equiv 1205731198732 equiv 1 (mod 119901
119894) 120573
ℎ
119894equiv 1 (mod 119901
119894)
1 le ℎ lt 1198732
119894 = 1 119904
(26)
Hence the congruence equation (22) has a total of 119904 2-dimensional NTTs mod 119901
119897119894
119894 119894 = 1 119904 From CRT and the
congruence equation (23) we deduce that the congruenceequation (24) holds
3 Two-Dimensional Overlap-Save Method
Select two positive integers 1198731015840 and 1198721015840 so that 1198731015840 + 119897
2minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 either 119873 or 119872 is given
by an integer power of 2 that is 119873 = 21198891 and 119872 = 2
1198892 where
1198891and 119889
2are two positive integers
Assuming that the input sequence 11990911989911198992
(1198991
=
0 1 1198971
minus 1 1198992
= 0 1 1198981
minus 1) is divided intomany small sections which are termed submatrices
The consecutive matrix is overlapped by the previous oneAs a simple example four blocks chosen from the dividedinput matrices are overlapped and are graphically shown inFigure 1 Parameters V
1and V
2can be considered as shift
parameters because changing either V1or V2will select dif-
ferent submatrix For the first submatrix shown in Figure 1(a)we select V
1= 0 and V
2= 0 Second submatrix (V
1= 0 and
V2= 1) is displayed in Figure 1(b)Third one with parameters
of V1
= 1 and V2
= 0 is given in Figure 1(c) Fourth matrixshown in Figure 1(d) has parameters V
1= 1 and V
2= 1 The
procedure of the overlapping is also graphically illustrated inFigure 1
For the operation of circular convolution it is necessaryto have two sequences with the same length Taking (7) for
example the sequences 11990911989911198992
and ℎ11989911198992
are of the same lengthof1198731times1198732 Since the input sequence has already divided into
submatrices of119873times119872 by (27) the filter sequence has to be oflength119873times119872Therefore119873minus119897
2augmenting zeros are required
to be added to the row and119872minus1198982zeros to the column of the
filter sequence ℎ11989911198992
As a result the sequence ℎ11989911198992
becomes
ℎ1015840
11989911198992
=ℎ11989911198992
if 1198991= 0 1 119897
2minus 1 119899
2= 0 1 119898
2minus 1
0 if 1198972le 1198991lt 119873 or 119898
2le 1198992lt 119872
(28)
where 1198991= 0 1 119873 minus 1 and 119899
2= 0 1 119872 minus 1
The following form shows how a sequence ℎ101584011989911198992
is yieldedby adding augmenting zeros to the filter sequence ℎ
11989911198992
ℎ 997904rArr ℎ1015840 =
ℎ
0 (29)
For each of the divided submatrices we can give V1and V
2
and compute the circular convolution of two 2D sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
The result of (31) is that of circular convolution As we willsee below there exists an important connection between thecircular convolution and the linear one In fact since 119899
1=
0 1 1198972minus 1 and 119899
2= 0 1 119898
2minus 1 we can choose 119905
1=
1198972minus1 119873minus1 and 119905
2= 1198982minus1 119872minus1 in (31) this makes
119873 gt 1199051minus 1198991
gt 0 and 119872 gt 1199052minus 1198992
gt 0 in (31) Thus theremainder can exactly be equal to 119905
In order to clearly understand the procedure of obtainingthe components of linear convolution from the circularconvolution equation (31) a graphic description of it can befound in Figure 2 We show a dashed rectangle with size of119873 by 119872 in Figure 2(a) to represent the circular convolutionwhich is described in (31) and a darkened rectangle with sizeof1198731015840 by119872
1015840 to indicate the linear convolution stated in (32)
6 Mathematical Problems in Engineering
2N
m
n
2M
M
N1 = 0 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N1 = 1 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
m
n
2N
2M
M
N
1 = 0 2 = 1
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(c)
m
n
2N
2M
M
N
1 = 1 2 = 1l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(d)
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
The consecutive matrix is overlapped by the previous oneAs a simple example four blocks chosen from the dividedinput matrices are overlapped and are graphically shown inFigure 1 Parameters V
1and V
2can be considered as shift
parameters because changing either V1or V2will select dif-
ferent submatrix For the first submatrix shown in Figure 1(a)we select V
1= 0 and V
2= 0 Second submatrix (V
1= 0 and
V2= 1) is displayed in Figure 1(b)Third one with parameters
of V1
= 1 and V2
= 0 is given in Figure 1(c) Fourth matrixshown in Figure 1(d) has parameters V
1= 1 and V
2= 1 The
procedure of the overlapping is also graphically illustrated inFigure 1
For the operation of circular convolution it is necessaryto have two sequences with the same length Taking (7) for
example the sequences 11990911989911198992
and ℎ11989911198992
are of the same lengthof1198731times1198732 Since the input sequence has already divided into
submatrices of119873times119872 by (27) the filter sequence has to be oflength119873times119872Therefore119873minus119897
2augmenting zeros are required
to be added to the row and119872minus1198982zeros to the column of the
filter sequence ℎ11989911198992
As a result the sequence ℎ11989911198992
becomes
ℎ1015840
11989911198992
=ℎ11989911198992
if 1198991= 0 1 119897
2minus 1 119899
2= 0 1 119898
2minus 1
0 if 1198972le 1198991lt 119873 or 119898
2le 1198992lt 119872
(28)
where 1198991= 0 1 119873 minus 1 and 119899
2= 0 1 119872 minus 1
The following form shows how a sequence ℎ101584011989911198992
is yieldedby adding augmenting zeros to the filter sequence ℎ
11989911198992
ℎ 997904rArr ℎ1015840 =
ℎ
0 (29)
For each of the divided submatrices we can give V1and V
2
and compute the circular convolution of two 2D sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
The result of (31) is that of circular convolution As we willsee below there exists an important connection between thecircular convolution and the linear one In fact since 119899
1=
0 1 1198972minus 1 and 119899
2= 0 1 119898
2minus 1 we can choose 119905
1=
1198972minus1 119873minus1 and 119905
2= 1198982minus1 119872minus1 in (31) this makes
119873 gt 1199051minus 1198991
gt 0 and 119872 gt 1199052minus 1198992
gt 0 in (31) Thus theremainder can exactly be equal to 119905
In order to clearly understand the procedure of obtainingthe components of linear convolution from the circularconvolution equation (31) a graphic description of it can befound in Figure 2 We show a dashed rectangle with size of119873 by 119872 in Figure 2(a) to represent the circular convolutionwhich is described in (31) and a darkened rectangle with sizeof1198731015840 by119872
1015840 to indicate the linear convolution stated in (32)
6 Mathematical Problems in Engineering
2N
m
n
2M
M
N1 = 0 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N1 = 1 2 = 0
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
m
n
2N
2M
M
N
1 = 0 2 = 1
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(c)
m
n
2N
2M
M
N
1 = 1 2 = 1l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(d)
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
Figure 1 Four divided input submatrices are overlapped using the 2D overlap-save method (a) the first submatrix with parameters V1
=
0 V2= 0 is located on the top-left corner (b) the second one (V
1= 1 V
2= 0) overlaps by the first submatrix and the size of the overlapping
between them is119873 times1198982minus1 (c) the third divided input submatrix has parameters of V
1= 0 V
2= 1 It is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119872 (d) the fourth one with V
1= 1 V
2= 1 is overlapped by the first submatrix
and the size of the overlapping between them is 1198972minus 1 times 119898
2minus 1
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(a)
m
n
2N
2M
M
N
l2 minus 1
m2 minus 1M
998400
N998400
2N998400
2M998400
(b)
Figure 2 Two consecutive output submatrices (a) the first block of V1= 0 and V
2= 0 and (b) the second block of V
1= 1 and V
2= 1
Mathematical Problems in Engineering 7
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
Obviously if the first 1198972minus 1 data in rows and the first 119898
2minus 1
data in columns are discarded from the dashed rectangle thedarkened rectangle can be obtained In the following we willexplain how to obtain the result of (4)
The output of the whole equation (4) can be representedin the following matrix
The corresponding graphic description is indicated inFigure 2(b) We will darken the rectangle in Figure 2(b) toexpress the submatrix (35) and the dashed one on its upper-left-corner to be the graphic description of the submatrix(34)
Clearly if (V1 V2) = (V10158401 V10158402) then (119905
1+ V11198731015840 1199052
+
V21198721015840) = (1199051015840
1+ V101584011198731015840 1199051015840
2+ V101584021198721015840) where 119897
2minus 1 le 119905
1 1199051015840
1le
119873 minus 1 1198982minus 1 le 119905
2and 1199051015840
2le 119872 minus 1 Let [119909] represent
the largest integer that is less than or equal to the realnumber 119909 therefore when V
1= 0 1 [(119897
1minus 119873)119873
1015840] and
V2
= 0 1 [(1198981minus 119872)119872
1015840] we will obtain all 1198731015840 times 119872
1015840
submatrices of thematrix119884 If 1198971= 11990411198731015840+119873 and119898
1= 11990421198721015840+
119872 then [(1198971minus 119873)119873
1015840] = (1198971minus 119873)119873
1015840 and [(1198981minus 119872)119872
1015840] =
(1198981minus119872)119872
1015840 The input matrix 11990911989911198992
will be divided with noremainder Figure 3 displays a series of small rectangles Eachrectangle stands for one1198731015840times119872
1015840 submatrixThe entirematrix119884 can be obtained by combining them
In Figure 3 there is a white Γ-type area indicating thatthere are no output data Two calculation methods canbe applied to obtain the data The first way is by using adirect method because only a few computations are neededAlternatively we can compute the data of Γ-type area with theoverlap-savemethod proposed above if1198731015840 = 119897
2minus1 and119872
1015840=
1198982minus 1 The detail of it is presented in the Technical Report
[18] The entire result is completed and is demonstrated inFigure 4 graphically
The two-dimensional overlap-save method is summa-rized as follows
m
n
Figure 3 All 1198731015840
times 1198721015840 submatrices of the output except the
boundary of Γ-type area
m
n
Figure 4 The full result of (4) computed by 2D overlap-savemethod where 119897
1= 11990411198731015840+ 119873 119898
1= 11990421198721015840+ 119872 and 119873
1015840+ 1198972minus 1 =
119873 lt 1198971and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1
(i) Select two positive integers 1198731015840 and 119872
1015840 so that 1198731015840 +1198972minus 1 = 119873 lt 119897
1and 119872
1015840+ 1198982minus 1 = 119872 lt 119898
1 Either
119873 or119872 is given by an integer power of 2 for using anFFT that is 119873 = 2
1198891 and 119872 = 2
1198892 where 119889
1and 119889
2
are two positive integers
(ii) Let ℎ101584011989911198992
satisfy (28)
(iii) Compute the circular convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 119873 minus 1 1198992
=
0 1 119872 minus 1) When 1199051
= 1198972minus 1 119873 minus 1 119905
2=
1198982minus 1 119872 minus 1 and V
1= 0 1 [(119897
1minus 119873)119873
1015840]
V2= 0 1 [(119898
1minus 119872)119872
1015840] we obtain all 1198731015840 times 119872
1015840
submatrices of the matrix of 119884 whose form is as (34)
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
overlap-save method The detail of it is presented inthe Technical Report [18]
We will introduce the way of using FNTT and 2Doverlap-save method with two examples One is numericalcomputation and the other is an experiment of extracting thecontour of a Chinese handwritten character
1198981= 11990421198721015840+ 119872 Suppose that 119886
1and 1198862satisfy 2
1198861minus1
lt 1198971+
1198972minus 1 le 2
1198861 = 119873
1and 1198861198862minus1
lt 1198981+1198982minus 1 le 2
1198862 = 119873
2 Let 119871
1
1198712 and 119871
3denote the numbers of multiplication necessary
for computing the convolution (4) by using direct methodFNTT or FFT to calculate (4) (using Proposition 1) and2D overlap-save method respectively In [18] we comparedthe number of multiplications necessary for computing theconvolution (4) of two 2D sequences 119909
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
119906 where 119906 ge 3 ℎ equiv 0 (mod 2)ℎ ge 4 then
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(45)
where 1198714denotes the number of multiplication necessary for
computing (4) by using 2D overlap-save method
Proof Because 1198731015840
= 1198972minus 1 1198721015840 = 119898
2minus 1 1198971= 1198981= 2119906+ℎ and
1198972= 1198982= 2119906+ 1 by (41) we have
119871 = (2119906+ 1)2sdot 2119906minus2
(2119906+ℎ+2
minus 3 (2119906+ 1) + 3)
= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
(46)
Compute the cyclic convolution of two sequences1199091198991+V111987310158401198992+V21198721015840 and ℎ
1015840
11989911198992
(1198991
= 0 1 2119906+2
minus 11198992
= 0 1 2119906+2
minus 1) that is compute matrix (35) byusing FNNT It is well known that using FNTT to calculatethe (35) requires
119873119872(3
2log119873119872 + 1) = 2
2(119906+2)(3119906 + 7) (47)
where 119873 = 119872 = 2119906+2
When 119873 = 119872 = 2119906+2 1198731015840 = 119872
1015840= 3 sdot 2
119906 119906 ge 3 ℎ equiv
0 (mod 2) and ℎ ge 4 then
[1198971minus 119873
1198731015840] = [
1198981minus 119872
1198721015840] = [
2119906+ℎ
minus 2119906+2
3 sdot 2119906]
= [
4 (2ℎminus2
minus 1)
3] =
4 (2ℎminus2
minus 1)
3
(48)
where 4(2ℎminus2
minus 1)3 is a positive integerSo the input matrix 119909
11989911198992
will be divided with noremainder When V
1= 0 1 [(119897
1minus 119873)119873
1015840] and V
2=
0 1 [(1198981minus119872)119872
1015840] we obtain all matrices as (35) Thus
1198714= (2119906+ 1)222(119906minus1)
(2ℎ+2
minus 3)
+ (4(2ℎminus2
minus 1)
3+ 1)
2
22(119906+2)
(3119906 + 7)
(49)
equal to (45) holds This completes the proof
Example 8 Suppose that 119906 = ℎ = 4 that is 1198971= 1198981= 256
1198972= 1198982= 17 119873 = 119872 = 64 and 119873
1015840= 1198721015840= 48 then
1198714= 3073856 (50)
If 1198971= 1198981= 256 119897
2= 1198982= 17 119873 = 119872 = 32 and 119873
1015840= 1198721015840=
16 then
1198713= 4194304 (51)
This example shows that 1198714lt 1198713in some particular cases
5 Experiments
The theoretic description of the 2D overlap-save method hasbeen provided in the preceding section This section furtherexamines how such a technique will be performed and howit will be applied to the wavelet transform Particularly wewill use image processing as an exampleThe experiment willfocus on the application of the 2D overlap-savemethod to thedyadic wavelet transform extracting the contours of Chinesehandwriting
51 Numerical Computation Let ℎ11989911198992
denote a digital filterand let a two-dimensional sequence
(ℎ11989911198992
) = (
1 0 1
1 1 0
0 1 1
) (52)
be the coefficients of the filter A two-dimensional sequence
(11990911989911198992
) =
((((
(
1 3 2 minus3 0 2 1 1
3 1 0 minus1 2 minus2 minus1 0
0 minus1 1 1 3 2 0 minus1
2 3 minus1 1 0 1 1 0
3 minus2 1 0 minus1 1 1 1
minus1 0 2 3 1 1 0 1
1 minus1 0 1 0 1 minus1 0
minus1 0 1 1 0 1 0 1
))))
)
(53)
10 Mathematical Problems in Engineering
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
represents the input of the filter In this example because 1198971=
1198981= 8 and 119897
2= 1198982= 3 the length of input matrix is 8 times 8
and that of the filter is 3 times 3Using the overlap-save method to calculate the output of
the filter
11991011989911198992
=
2
sum
1198961=0
2
sum
1198962=0
ℎ11989611198962
1199091198991minus11989611198992minus1198962
0le1198991minus1198961lt8
0le1198992minus1198962lt8
1198991= 0 1 2 3 4 5 6 7 119899
2= 0 1 2 3 4 5 6 7
(54)
means that the elements of the matrix
119884 = (
11991000
11991001
sdot sdot sdot 11991007
11991010
11991011
sdot sdot sdot 11991017
11991070
11991071
sdot sdot sdot 11991077
) (55)
need to be computedThe solution is given belowSince 119897
1= 1198981= 8 1198972= 1198982= 3 by selecting 119873
1015840= 1198972minus 1 =
2 and 1198721015840= 1198982minus 1 = 2 we have 119873 = 119872 = 4 In order to
use the FNTT the input matrix is divided into 9 submatricesof size 4times 4 Each consecutive submatrix is partly overlappedwith the size of 2 times 2 pixels In addition zeros of size 1 times 1
are added to the filtering sequence to ensure that its lengthis equivalent to that of the submatrix Therefore the filteringsequence ℎ
11989911198992
is extended to ℎ1015840
11989911198992
that is
ℎ1015840
11989911198992
= ℎ11989911198992
if 1198991= 0 1 2 119899
2= 0 1 2
0 if 1198991= 3 or 119899
2= 3
(56)
where 1198991= 0 1 2 3 119899
2= 0 1 2 3
Next we compute the circular convolution of the twosequences 119909
where 4minus1 denotes an integer so that (4minus1) sdot 4 equiv 1 (mod 257)
Mathematical Problems in Engineering 11
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
linear convolutionof the output matrixof the output matrix
cyclic convolutionCalculate the
of the filtering matrix
Calculate the
Calculate the numbertheoretic transformof the input matrix
Choose M
theoretic transformCalculate the number
Figure 5 Five steps of applying FNTT to calculate the output of (54)
Step 2 (calculating the number theoretic transform of inputmatrix by the FNTT) For computational convenience hereone writes
11988001198962
equiv
3
sum
1198992=0
11990901198992
1611989921198962 119880
11198962
equiv
3
sum
1198992=0
11990911198992
1611989921198962(mod 257)
11988021198962
equiv
3
sum
1198992=0
11990921198992
1611989921198962 119880
31198962
equiv
3
sum
1198992=0
11990931198992
1611989921198962(mod 257)
1198962= 0 1 2 3
(64)
Then (62) becomes
11988311989610
equiv
3
sum
1198991=0
11988011989910
1611989911198961 119883
11989611
equiv
3
sum
1198991=0
11988011989911
1611989911198961(mod 257)
11988311989612
equiv
3
sum
1198991=0
11988011989912
1611989911198961 119883
11989613
equiv
3
sum
1198991=0
11988011989913
1611989911198961(mod 257)
1198961= 0 1 2 3
(65)
Use the FNTT to calculate (64) and (65)So
(
11988000
11988001
11988002
11988003
11988010
11988011
11988012
11988013
11988020
11988021
11988022
11988023
11988030
11988031
11988032
11988033
)
equiv (
3 95 3 minus97
3 35 3 minus29
1 minus33 1 31
5 35 minus3 minus29
) (mod 257)
119883 = (
11988300
11988301
11988302
11988303
11988310
11988311
11988312
11988313
11988320
11988321
11988322
11988323
11988330
11988331
11988332
11988333
)
equiv (
12 132 4 minus124
minus30 128 98 minus128
minus4 minus8 4 minus8
34 128 minus94 minus128
) (mod 257)
(66)
Step 3 (calculating the number theoretic transformof the filterby the FNTT) The FNTT of the filter ℎ
11989911198992
is written as
11986711989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
ℎ1015840
11989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(67)
Similarly we have
119867 = (
11986700
11986701
11986702
11986703
11986710
11986711
11986712
11986713
11986720
11986721
11986722
11986723
11986730
11986731
11986732
11986733
)
equiv (
6 32 2 minus32
32 0 2 34
2 minus2 2 minus2
minus32 minus30 2 0
) (mod 257)
(68)
Step 4 (compute the number theoretic transformof the outputof the filter) From the preceding three steps we can computethe circular convolution of the filter by defining the productof 119883 and 119867 as
119884 = (
11988400
11988401
11988402
11988403
11988410
11988411
11988412
11988413
11988420
11988421
11988422
11988423
11988430
11988431
11988432
11988433
)
equiv 119883 ⊙ 119867 = (
11988300
11986700
11988301
11986701
11988302
11986702
11988303
11986703
11988310
11986710
11988311
11986711
11988312
11986712
11988313
11986713
11988320
11986720
11988321
11986721
11988322
11986722
11988323
11986723
11988330
11986730
11988331
11986731
11988332
11986732
11988333
11986733
)
equiv (
72 112 8 113
68 0 minus61 17
minus8 16 8 16
minus60 15 69 0
) (mod 257)
(69)
where
11988411989611198962
equiv
3
sum
1198991=0
3
sum
1198992=0
11991011989911198992
1611989911198961 sdot 1611989921198962(mod 257)
1198961= 0 1 2 3 119896
2= 0 1 2 3
(70)
12 Mathematical Problems in Engineering
Table 1 Filtering coefficients 12059521198951
119896119897 (119895 = 2) 120595(119903) are a quadratic spline wavelet
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
Figure 6 A 2-dimensional image of an aircraft was processed with 2D overlap-save method (a) Original Chinese handwriting and (b) thedivided up subimages of the character
and the 11991011989911198992
(1198991
= 0 1 2 3 1198992
= 0 1 2 3) is the circularconvolution of (60)
Step 5 (calculate the output of the linear convolution) From(63) we deduce the inverse transform of (70)
11991011989911198992
equiv 4minus1
sdot 4minus1
3
sum
1198961=0
3
sum
1198962=0
11988411989611198962(minus16)11989911198961
sdot (minus16)11989921198962(mod 257)
1198991= 0 1 2 3 119899
2= 0 1 2 3
(71)
Using FNTT to calculate (71) we have
(
11991000
11991001
11991002
11991003
11991010
11991011
11991012
11991013
11991020
11991021
11991022
11991023
11991030
11991031
11991032
11991033
) equiv (
8 6 4 0
1 7 13 1
2 2 6 4
1 5 5 7
) (mod 257)
(72)
Then the linear convolution in (58) is obtained
(11991022
11991023
11991032
11991033
) = (11991022
11991023
11991032
11991033
) = (6 4
5 7) (73)
52 Application to Image Processing The proposed methodcan be applied to many fields such as image processingand pattern recognition An example of the application to
the image processing is illustrated in Figure 6 The originalimage shown in Figure 6(a) is a 2-dimensional image of anaircraft with a size of 256 times 256 pixels that is 119897
1= 1198981
=
256 Our task is to extract its contours The discrete wavelettransform with scale 119904 = 2
119895 where 119895 = 2 was applied and thespline wavelet described in (1) of size 17 times 17 was chosenthat is 119897
2= 1198982
= 17 We use the filtering coefficientstabulated in Table 1 Using the 2D overlap-save method thisoriginal image was divided into 25 sections with each onehaving the dimension of 64 times 64 pixels that is we select119873 = 119872 = 64 1198731015840 = 119872
1015840= 48 [(119897
1minus 119873)119873
1015840] = [(119898
1minus
119872)1198721015840] = 4 (See Example 8 in Section 4) Figure 6(b)
shows these subimages The overlapped part can be explicitlyobserved in each subimage
To facilitate the presentation of the 2D overlap-savemethod the procedure of it is diagonally displayed inFigure 7 where a series of the divided subimages are illus-tratedThe image in Figure 6(a) was equivalently divided into25 sections by applying (27) In Figure 7(a) we represent 5subimages of V
0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2
and V0
= 3 V1
= 3 and V0
= 4 V1
= 4 consecutively Thelower-right corner of each subimage is repeated in the upper-left corner of its consecutive subimage Four subimages aredisplayed in Figure 7(b) and the parameters of V
0= 0 V1= 1
V0
= 1 V1
= 2 V0
= 2 V1
= 3 and V0
= 3 V1
= 4 werechosen In Figure 7(c) we selected 3 subimages of V
0= 0
V1
= 2 V0
= 1 V1
= 3 and V0
= 2 V1
= 4 There are3 subimages of V
0= 0 V
1= 3 and V
0= 1 V
1= 4 in
Figure 7(d) Figure 7(e) displays subimages of V0= 0 V
1= 4
Mathematical Problems in Engineering 13
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
Figure 7 A 2-dimensional image of an aircraft is dividing into separated small subimages the parameters of V0and V
1are chosen as follows
(a) V0= 0 V
1= 0 V
0= 1 V
1= 1 V
0= 2 V
1= 2 and V
0= 3 V
1= 3 and V
0= 4 and V
1= 4 (b) V
0= 0 V
1= 1 V
0= 1 V
1= 2 V
0= 2 V
1= 3 and
V0= 3 and V
1= 4 (c) V
0= 0 V
1= 2 V
0= 1 V
1= 3 and V
0= 2 and V
1= 4 (d) V
0= 0 V
1= 3 and V
0= 1 and V
1= 4 (e) V
0= 0 and V
1= 4 (f)
V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2 and V
0= 4 and V
1= 3 (g) V
0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4 and V
1= 2 (h) V
0= 3 V
1= 0
and V0= 4 and V
1= 1 and (i) V
0= 4 and V
1= 0
In Figure 7(f) V0= 1 V
1= 0 V
0= 2 V
1= 1 V
0= 3 V
1= 2
and V0= 4 V1= 3were chosen In Figure 7(g) the parameters
were decided upon V0= 2 V
1= 0 V
0= 3 V
1= 1 and V
0= 4
V1= 2 Two subimages are shown in Figure 7(h) by selecting
V0= 3 V
1= 0 and V
0= 4 V
1= 1 There is one subimage of
V0= 4 V
1= 0 in Figure 7(i)
We select a 2D NTT
11988311989611198962
equiv
63
sum
1198991=0
63
sum
1198992=0
11990911989911198992
211989911198961211989921198962 (mod 2
32+ 1) (74)
14 Mathematical Problems in Engineering
(a) (b)
(c) (d)
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
Figure 8 The linear outputs of 2D overlap-save method (a) the outputs of the linear convolution of each subimages (b) the left border andthe upper border of the outputs of the image of aircraft (c) the combined outputs of linear convolution except Γ-type border and (d) theentire contour extracted using 2D overlap-save method under wavelet transform
which has inverse transform
11990911989911198992
equiv 64minus1
64minus1
63
sum
1198961=0
63
sum
1198962=0
11988311989611198962
(minus231)11989911198961
times (minus231)11989921198962
(mod 232
+ 1)
1198991= 0 1 63 119899
2= 0 1 63
(75)
where 64minus1 denotes integer so that (64minus1)64 equiv 1 (mod 2
32+
1) and |11990911989911198992
| le 255The FNTT was applied to calculate the circular convo-
lution of each subimage After applying (31) to each of thesubblocks a series of segments of the contours were obtained
By using (32) we can gradually reach the final result of thelinear convolution Figure 8(a) illustrates those pixels whichsatisfy (35) where 119897
= 0 1 2 3 4 Combining them together thecontours of the whole character were obtained and displayedin Figure 8(c) which excludes the Γ-type border Next the
Γ-type border was computed by direct method Figure 8(b)shows the result Combining this result with that shown inFigure 8(c) the entire contours of the Chinese handwritingwere combined Figure 8(d) represents the entire contour
6 Conclusions
A novel approach to reduce the computation complexityof the wavelet transform has been presented in this paperTwo key techniques have been applied in this new methodnamely fast number theoretic transform (FNTT) and two-dimensional overlap-save technique In the fast numbertheoretic transform the linear convolution is replaced bythe circular convolution It can speed up the computationof 2D discrete wavelet transform Directly calculating thefast number theoretic transform to the whole input sequencemay meet two difficulties namely a big modulo obstructsthe effective implementation of the fast number theoretictransform and a long input sequence slows the computation
Mathematical Problems in Engineering 15
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998
of the fast number theoretic transform down To fight withsuch deficiencies a new technique which is referred to as 2Doverlap-save method has been developed Experiments havebeen conducted The fast number theoretic transform and2D overlap-method have been used to implement the dyadicwavelet transform and applied to contour extraction in imageprocessing and pattern recognition
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgments
This work was supported by the Research GrantsMYRG205(Y1-L4)-FST11-TYY and MYRG187(Y1-L3)-FST11-TYY and Chair Prof Grants RDG009FST-TYY ofUniversity of Macau as well as Macau FDC Grants T-100-2012-A3 and 026ndash2013-A This research project was alsosupported by the National Natural Science Foundation ofChina 61273244
References
[1] P L Combettes and J-C Pesquet ldquoWavelet-constrained imagerestorationrdquo International Journal of Wavelets Multiresolutionand Information Processing vol 2 no 4 pp 371ndash389 2004
[2] M Ehler and K Koch ldquoThe construction of multiwavelet bi-frames and applications to variational image denoisingrdquo Inter-national Journal of Wavelets Multiresolution and InformationProcessing vol 8 no 3 pp 431ndash455 2010
[3] P Jain and S N Merchant ldquoWavelet-based multiresolutionhistogram for fast image retrievalrdquo International Journal ofWavelets Multiresolution and Information Processing vol 2 no1 pp 59ndash73 2004
[4] S Mallat and W L Hwang ldquoSingularity detection and pro-cessing with waveletsrdquo Institute of Electrical and ElectronicsEngineers Transactions on InformationTheory vol 38 no 2 pp617ndash643 1992
[5] B U Shankar S K Meher and A Ghosh ldquoNeuro-waveletclassifier formultispectral remote sensing imagesrdquo InternationalJournal of Wavelets Multiresolution and Information Processingvol 5 no 4 pp 589ndash611 2007
[6] S Shi Y Zhang and Y Hu ldquoA wavelet-based image edgedetection and estimationmethod with adaptive scale selectionrdquoInternational Journal of Wavelets Multiresolution and Informa-tion Processing vol 8 no 3 pp 385ndash405 2010
[7] Y Y Tang Wavelets Theory Approach to Pattern RecognitionWorld Scientific Singapor 2009
[8] Q M Tieng and W Boles ldquoWavelet-based affine invariantrepresentation a tool for recognizing planar objects in 3Dspacerdquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 19 no 8 pp 921ndash925 1997
[9] P Wunsch and A F Laine ldquoWavelet descriptors for mul-tiresolution recognition of handprinted charactersrdquo PatternRecognition vol 28 no 8 pp 1237ndash1249 1995
[10] H Yin and H Liu ldquoA Bregman iterative regularization methodfor wavelet-based image deblurringrdquo International Journal of
Wavelets Multiresolution and Information Processing vol 8 no3 pp 485ndash499 2010
[11] X You and Y Y Tang ldquoWavelet-based approach to characterskeletonrdquo IEEE Transactions on Image Processing vol 16 no 5pp 1220ndash1231 2007
[12] T Zhang Q Fan and Q Gao ldquoWavelet characterization ofHardy space ℎ
1 and its application in variational image decom-positionrdquo International Journal of Wavelets Multiresolution andInformation Processing vol 8 no 1 pp 71ndash87 2010
[13] Y Y Tang L Y Li Feng and J Liu ldquoScale-independent waveletalgorithm for detecting step-structure edgesrdquo Tech Rep IEEETransactions on InformationTheory Brisbane Australia 1998
[14] S G Mallat ldquoMultiresolution approximations and waveletorthonormal bases of 119871
2(R)rdquo Transactions of the American
Mathematical Society vol 315 no 1 pp 69ndash87 1989[15] J H McClellan and C M Rader Number Theory in Digital Sig-
nal Processing Prentice Hall Signal Processing Series PrenticeHall Englewood Cliffs NJ USA 1979
[16] X Ran and K J R Liu ldquoFast algorithms for 2-D circularconvolutions and number theoretic transforms based on poly-nomial transforms over finite ringsrdquo IEEE Transactions onSignal Processing vol 43 no 3 pp 569ndash578 1995
[17] Q Sun ldquoSome results in the application of the number theoryto digital signal processing and public-key systemsrdquo in NumberTheory and Its Applications in China vol 77 of ContemporaryMathematics pp 107ndash112 American Mathematical SocietyProvidence RI USA 1988
[18] Y Y Tang Q Sun L H Yang and L Feng ldquoAn overlap-save method for the calculation of the two-dimensional digitalconvolutionrdquo Tech Rep Department of Computer ScienceHong Kong Baptist University 1998