Fourier Analysis in Polar and Spherical.pdf

7/27/2019 Fourier Analysis in Polar and Spherical.pdf

1/26

ALBERT-LUDWIGS-UNIVERSIT AT FREIBURGINSTITUT F UR INFORMATIK

Lehrstuhl f ur Mustererkennung und Bildverarbeitung

Fourier Analysis in Polar and SphericalCoordinates

Internal Report 1/08

Qing Wang, Olaf Ronneberger, Hans Burkhardt


2/26

Fourier Analysis in Polar and SphericalCoordinates

Qing Wang, Olaf Ronneberger, Hans Burkhardt

Abstract

In this paper, polar and spherical Fourier Analysis are dened as thedecomposition of a function in terms of eigenfunctions of the Laplacianwith the eigenfunctions being separable in the corresponding coordinates.Each eigenfunction represents a basic pattern with the wavenumber in-dicating the scale. The proposed transforms provide an effective radialdecomposition in addition to the well-known angular decomposition. Thederivation of the basis functions is compactly presented with an emphasison the analogy to the normal Fourier transform. The relation betweenthe polar or spherical Fourier transform and normal Fourier transform isexplored. Possible applications of the proposed transforms are discussed.

1 Introduction

Fourier transform is very important in image processing and pattern recognition

both as a theory and as a tool. Usually it is formulated in Cartesian coordinates,where a separable basis function in 3D space without normalization is

ei k r = eik x x eik y y eik z z (1)

where (x,y,z ) are coordinates of the position r and kx , ky , kz are componentsof the wave vector k along the corresponding axis. The basis function (1)represents a plane wave. Fourier analysis is therefore the decomposition of afunction into plane waves. As the basis function is separable in x, y and z, Thedecomposition can be understood as being made up of three decompositions (for3D).

The Laplacian is an important operator in mathematics and physics. Itseigenvalue problem gives the time-independent wave equation. In Cartesian

coordinates the operator is written as

2 =

2x +

2y +

2z =

2

x 2+

2

y 2+

2

z 2.

for 3D space. (1) is an eigenfunction of the Laplacian and is separable in Carte-sian coordinates.

When dened on the whole space, functions given in (1) are mutually or-thogonal for different k ; wave vectors take continuous values and it is said thatone has a continuous spectrum. Over nite regions, the mutual orthogonalitygenerally does not hold. To get an orthogonal basis, k can only take values from

1


3/26

a discrete set and the spectrum becomes discrete. The continuous Fourier trans-

form reduced to Fourier series expansion (with continuous spatial coordinates )or to the discrete Fourier transform (with discrete spatial coordinates).

For objects with certain rotational symmetry, it is more effective for them to beinvestigated in polar (2D) or spherical (3D) coordinates. It would be of greatadvantage if the image can be decomposed into wave-like basic patterns thathave simple radial and angular structures, so that the decomposition is madeup of radial and angular decompositions. Ideally this decomposition should bean extension of the normal Fourier analysis and can therefore be called Fourieranalysis in the corresponding coordinates. To fulll these requirements, thebasis functions should take the separation-of-variable form:

R(r )() (2)

for 2D andR(r )()() = R(r ) (,) (3)

for 3D where ( r, ) and ( r, , ) are the polar and spherical coordinates respec-tively. They should also be the eigenfunctions of the Laplacian so that theyrepresent wave-like patterns and that the associated transform is closely relatedto the normal Fourier transform. The concrete form of the angular and radialparts of the basis functions will be investigated and elaborated in the comingsections but will be briey introduced below in order to show previous workrelated to them.

For polar coordinates, as will be shown in the next section, the angular part

of a basis function is simply

() =1

2 eim (4)

where m is an integer, which is a natural result of the single-value requirement:() = ( + 2 ), a special kind of boundary condition. The associated trans-form in angular coordinate is nothing else but the normal 1D Fourier transform.For spherical coordinates, the angular part of a basis function is a spherical har-monic

(,) = Y lm (,) = 2l + 14 (l m)!(l + m)!P lm (cos )eim (5)where P lm is an associated Legendre polynomial and l and m are integers, l

0

and |m| l. It also satises the single-value requirement. The correspondingtransform is called Spherical Harmonic (SH) transform and has been widelyused in representation and registration of 3D shapes [810].

The angular parts of the transforms in 2D and 3D are therefore very familiar.Not so well-known are the transforms in the radial direction. The radial basisfunction is a Bessel function J m (kr ) for polar coordinates and a spherical Besselfunction j l (kr ) for spherical coordinates. In both cases, The parameter k cantake either continuous or discrete values, depending on whether the region isinnite or nite. For functions dened on (0 , ), the transform with J m (kr ) asintegral kernel and r as weight is known as the Hankel transform. For functions

2


4/26

dened on a nite interval, with zero-value boundary condition for the basis

functions, one gets the Fourier-Bessel series [1]. Although the theory on Fourier-Bessel series has long been available, it mainly has applications in physics-relatedareas [18, 19]. [12] and a few references therein are the only we can nd thatemploy Fourier-Bessel series expansion for 2D image analysis. Methods basedon Zernike moments are on the other hand much more popular in applicationswhere we believe the Fourier-Bessel expansion also ts. The Zernike polynomialsare a set of orthogonal polynomials dened on a unit disk, which have the sameangular part as (4).

The SH transform works on the spherical surface. When it is used for 3Dvolume data, the SH features (extracted from SH coefficients) can be calculatedon concentric spherical surfaces of different radii and be collected to describean object, as suggested in [9]. This approach treats each spherical surface asindependent to one another and has a good localization nature. it fails to de-scribe the relation of angular properties of different radius as a whole, thereforecannot represent the radial structures effectively. The consideration of how todescribe the radial variation of the SH coefficients actually motivated the wholework presented here.

In this paper, the operations that transform a function into the coefficients of the basis functions given in (2) and (3) and described above will simply be calledpolar and spherical Fourier transform respectively. It should be noted thoughthat in the literature, the former often refers to the normal Fourier transformwith wave vectors k expressed in polar coordinates ( k,k ) [16] and the latteroften refers to the SH transform [17].

Due to the extreme importance of the Laplacian in physics, the expansion

of functions with respect to its eigenfunctions is naturally not new there. Forexample, in [20] and [21], the eigenfunctions of the Laplacian are used for expan-sion of sought wave functions. The idea that these eigenfunctions can be usedas basis functions for analyzing 2D or 3D images is unfamiliar to the patternrecognition society. There also lacks a simple and systematic presentation of theexpansion from the point of view of signal analysis. Therefore, although partsof the derivation are scattered in books like [1], we rederive the basis functionsto emphasize the analogy to the normal Fourier transform. Employment of the Sturm-Liouville theory makes this analogy clearer and the derivation morecompact.

The proposed polar and spherical Fourier transforms are connected with thenormal Fourier transform by the Laplacian. We investigate the relations betweenthem so that one can understand the proposed transforms more completely anddeeply. It is found that the relations also provide computational convenience.An advantage of the proposed transforms is that when a function is rotatedaround the origin, the change of its transform coefficients can be relativelysimply expressed in terms of the rotation parameters. This property can, onthe one hand, be used to estimate rotation parameters, on the other hand, beused to extract rotation-invariant descriptors. We will show how to do them.

Section 2 deals with the polar Fourier transform. Besides presentation of the theory, issues about calculation of the coefficients are discussed. A shortcomparison between polar Fourier basis functions and Zernike functions is madeat the end. Parallel to section 2, the theory for the spherical Fourier transformis given in section 3. In section 4 we investigate the possible applications of the

3


5/26

polar and spherical Fourier transforms. At the end, conclusion and outlook are

given.

2 Polar Fourier transform

2.1 Basis Functions

2.1.1 Helmholtz Equation and Angular Basis Functions

As a direct extension from the Cartesian case, we begin with the eigenfunctionsof the Laplacian, whose expression in polar coordinates is given by:

2 =

2r +

1r 2

2 (6)

where

2r =

1r

r

r r

(7)

and

2 =

2

2. (8)

are the radial and angular parts. The eigenvalue problem can be written as

2r (r, ) +

1r 2

2(r, ) + k

2(r, ) = 0 , (9)

which is the Helmholtz differential equation in polar coordinates. We require

that k2

0 as with negative k2

, the radial functions are exponentially growingor decaying, which are not interesting for our purpose. It will be shown later thatsuch a requirement does not prevent the eigenfunctions from forming a basis.For simplicity, it is further required that k 0. Substituting the separation-of-variable form ( r, ) = R(r )() into (9), one gets

2

2 + m2 = 0 (10)

1r

r

r r

R + k2 m2

r 2R = 0 . (11)

The solution to (10) is simply

m () = 1 2 eim (12)

with m being an integer.

2.1.2 Radial Basis Functions

The general solution to (11) is

R(r ) = AJ m (kr ) + BY m (kr ) (13)

4


6/26

where J m and Y m are the m-th order Bessel functions and Neumann functions

respectively [1]; A and B are constant multipliers. A nonsingular requirementof R at the origin leavesR(r ) = J m (kr ) (14)

as Y m is singular at the origin. Bessel functions satisfy the orthogonality relation

0 J m (k1r )J m (k2r )rdr = 1k1 (k1 k2) (15) just like the complex exponential functions satisfy

eik 1 x eik 2 x dx = 2 (k1 k2) . (16)Actually J m (kr ) forms a basis for functions dened on (0 , ) and satisedcertain continuous and integrable conditions (Later this kind of description isunderstood when we talk about functions to be transformed or to be expanded).

For the Fourier transform, an innite space corresponds to a continuousspectrum and a nite space corresponds to a discrete spectrum, where properboundary conditions select the spectrum. The same is also true for the radialbasis functions in polar coordinates. Over the nite interval [0 , a], the orthogonalrelation like in (15) generally does not hold any more, instead,

a0 J m (k1r )J m (k2r )rdr=

ak21

k22

[k2J m (k1a)J m (k2a) k1J m (k2a)J m (k1a)] . (17)

By imposing boundary conditions according to the Sturm-Liouville (S-L) theory[2, 5], a set of k values can be determined that make J m (kr ) again mutuallyorthogonal. We rst rewrite (11) as

(rR ) +m2

rR = k2rR. (18)

With p(r ) = rq (r ) = m

2

rw(r ) = r = k2

, (19)

the equation (18) takes the S-L form:

( p(r )R ) + q (r )R = w(r )R (20)

where r[0, a]. Eq.(20) together with the following boundary conditions formsa S-L system.R(0)cos p(0)R (0) sin = 0R(a)cos p(a)R (a)sin = 0

(21)

where , [0, ). The allowed values of are called the eigenvalues of thesystem. According to the theory, for such a S-L system,

5


7/26

1. The eigenvalues are nonnegative real numbers and can be numbered to

form an increasing sequence 1 < 2 < < n < ;2. The corresponding eigenfunctions can be uniquely determined up to acostant multiplier;

3. The eigenfunctions are mutually orthogonal with respect to the weightfunction w(r ) = r ;

4. The n-th eigenfunction has exactly n 1 zeros on the interval (0 , a );5. The complete set of eigenfunctions forms a complete orthogonal set of

functions dened on the interval [0 , a ].

Since R(r ) = J m (kr ) is a general non-singular solution to (20), the values

k can take (therefore the eigenvalues = k2

) are determined by the boundaryconditions (21). With = / 2, the rst equation in (21) has actually no effecton the selection of k but Y m can be excluded from the general solution (13) if we have not done so. The only effective boundary condition left is the secondequation in (21). Substituting R(r ) = J m (kr ) into it, one gets

J m (ka )cos kaJ m (ka )sin = 0 (22)with x = ka , (22) becomes

J m (x)cos xJ m (x)sin = 0 (23)Suppose ( xm 1 < x m 2 < < x mn < ) are nonnegative solutions to (23) withJ

m(x

mnr/a ) being nonzero functions, then k can take the values from

xm 1a

,xm 2

a, ,

xmna

, .Dene

knm =xmn

a(24)

(The indices n and m now exchange their order for the sake of convention), then-th eigenvalue is then n = k2nm and the n-th eigenfunction is J m (knm r ). Theorthogonality of the eigenfunctions can be written as

a0 J m (knm r )J m (kn m r )rdr = N (m )n nn . (25)By taking the limit of (17) as k2 k1 and taking into account that J m (kr ) isthe solution to (11), one can get

N (m )n =a2

2J m

2(xmn ) + 1 m2

x2mnJ 2m (xmn ) . (26)

The normalized radial function can therefore be dened as

Rnm (r ) =1

N (m )n J m (knm r ) . (27)6


8/26

{Rnm |n = 1 , 2, }forms an orthonormal basis on the interval [0 , a ]. A functionf (r ) dened on this interval can be expanded as

f (r ) =

n =1 a0 f ()Rnm ()d Rnm (r ) . (28)So far in (23) has not been specied. Two cases are interesting:

Zero-value boundary condition: with sin = 0, (23) reduces to

J m (x) = 0 . (29)

(note that x = ka ). xmn should be the positive zeros of J m (x). Under thiscondition,

N (m )n

=a2

2J 2

m +1(x

mn) (30)

and the right-hand side of (28) is usually known as m-th order Fourier-Besselseries of f (r ).Derivative boundary condition: with cos = 0, (23) becomes

J m (x) = 0 . (31)

xmn should be the zeros of J m (x). One special case needs to be considered here:x = 0 is one solution to J 0(x) = 0 and J 0(0 r/a ) = 1 has exactly 0 zero on(0, a). According to the S-L theory, x = 0 should be recognized as x01 . Underthis boundary condition,

N (m )n =

a2

2 1 m2

x2mn J 2m (xmn ) (32)

with the special case N (0)1 = a2 / 2.It is clear now that different boundary conditions lead to different spectra

of the system. The choice should depend on the problems under investigation.To give an impression how the radial functions look like, we show the rst fewof them for m = 2 with the zero and the derivative boundary conditions in Fig.1 (a) and (b). It is intuitive to choose the zero boundary condition when theimages tend to be zero at r = a and the derivative condition when the imagetend to be constant in radial direction near r = a. Often it is necessary to dosome experiments to nd the better choice.

Rnm (r ) has n

1 zeros on (0, a). Its wave-like property can be made more

clear by considering the asymptotic behavior of the Bessel functions [1]. Onehas

Rnm (r )1 r cos knm r

m2

4

(33)

for knm r |m2 14 |. Therefore, with large knm r , Rnm (r ) approaches a cosinefunction with its amplitude decreasing as fast as 1 / r . There is a phase shift of (m/ 2 + / 4), which is corresponded to by a delay of the function to takethe wave-like form near the origin. See Fig. 6 (a) for a typical form of Rnm (r )with relatively large n and m.

7


9/26

0 0.2 0.4 0.6 0.8 1

1

0.5

0

0.5

1

1.5

2

r

1

2

34

(a)

0 0.2 0.4 0.6 0.8 1

1

0.5

0

0.5

1

1.5

2

r

1

2

3

4

(b)

0 0.2 0.4 0.6 0.8 1

1.5

1

0.5

0

0.5

1

1.5

2

2.5

3

r

2

6

48

(c)

Figure 1: The rst few radial basis functions for 2D with m = 2 and a = 1: (a)Rnm with zero boundary condition; (b) Rnm with derivative boundary conditionand (c) normalized radial Zernike function Z nm . The number beside each curveis the value of n.

8


10/26

2.1.3 Basis Functions

The basis function for the polar Fourier transform is composed of the radial andthe angular parts. Consequently, for the transform dened on the whole space,the basis function is given by

k,m (r, ) = kJ m (kr )m () (34)with k taking continuous nonnegative values and m dened by (12). For thetransform dened on the nite region r a, the basis function is given by

nm (r, ) = Rnm (r )m () (35)

with Rnm dened by (27). The orthogonality relation is given by:

a

0 2

0

nm (r, )n m (r, )rdrd = nn mm . (36)

A basis function satises the following equation as well as the correspondingboundary conditions

2nm + k2nm nm = 0 . (37)

{nm }with ( n = 1 , 2, ) and ( m = , 2, 1, 0, 1, 2, ) form an orthonor-mal basis on the region r a.For nm (r, ), m is the number of periods in the angular direction, andn 1 corresponds to the number of zero crossings in the radial direction. Asfor the meaning of knm , those who are familiar with quantum mechanics canrecognize from (37) that k2nm is the energy level (except for a constant factor) of the system and its corresponding wave function is nm . Some of the functions

with lowest energy levels are shown in Figure 2. One can nd that the higherthe energy level, the ner the structures. Therefore for image analysis, the valueof k is an indication of the scale of the basic patterns, which is consistent withthe normal Fourier transform.

2.2 Expansion

A 2D function f (r, ) dened on the whole space can be expanded with respectto k,m as dened in (34):

f (r, ) = 0 m = P k,m k,m (r, ) kdk (38)where

P k,m = 0 20 f (r, )k,m (r, )rdrd (39)are the polar Fourier coefficients ( P stands for P olar). The innite transform asgiven in (38) and (39) is mainly of theoretical interest. In practice, one shoulduse the transform dened on a nite region. A function f (r, ) dened on r acan be expanded with respect to {nm }(It is understood with this symbol thatn and m are integers and n is positive) as

f (r, ) =

n =1

m = P nm nm (r, ) (40)

9


11/26

(1,0) (1,1) (1,2)

(2,0) (1,3) (2,1)

(1,4) (2,2) (3,0)

(1,5) (2,3) (1,6)

Figure 2: Basic patterns represented by nm with zero boundary condition.Shown are the real part of the functions. ( n, m ) pairs are given under eachpattern. The patterns are listed in the increasing order of the value of knm .

10


12/26

m

n

0 10 20 30 40

10

20

30

Figure 3: Isolines of knm

where the coefficients

P nm = a0 20 f (r, )nm (r, )rdrd . (41)There are two indices for the expansion. How should the terms be ordered andtherefore be truncated for a nite-term expansion? A natural way is accordingto the energy levels. In the language of image analysis, according to the scalesof the basic patterns. Larger-scale patterns should be taken into account rst.This is often the best choice if no other information about the data is available.Figure 3 shows the isolines of knm .

A digital image is usually given on an equally-spaced grid in Cartesian coor-dinates. To evaluate the coefficients as in (41), it is advisable to map the image I into polar coordinates, where the transform becomes separable and the angularpart can be done fast with FFT. The grid density of the mapped image I polarshould be high enough to accommodate the nest patterns in the expansion.Let the largest values for m and knm be mmax and kmax . Denote the radial andangular size of I polar as M r and M . The sampling theorem requires

M 2mmax . (42)M should also be chosen to facilitate fast calculations.

Considering the asymptotic behavior of Rnm (r ) in (33), knm takes the posi-tion of the wavenumber. One can expect that M r should be at least

2 a

2/k max=

akmax

.

However, the right-hand side of (33) is only the asymptotic bahavior and is nota real trigonometric function. Furthermore, the weight function is r instead of 1 for the radial integral. It is necessary to do some numerical experiments todetermine the number of sampling points M r in order to ensure that (25) holdswithin a certain relative error.

Mapping of the image is a process of interpolating and sampling. It mustbe handled carefully to avoid aliasing. The nest structure supported in both

11


13/26

radial and angular directions should match the nest structure in Cartesian

coordinates. Approximately, that means, for a disk of radius a (in the unit of apixel) in the original image, there should be 2a steps in r and 2 2a steps in. The 2 inside the expressions comes from the fact than the highest frequencyfor the original image is 2/ 2 instead of 1/ 2. Often 2 can be dropped if oneis sure that there is no so ne structure in the original image, which is usuallyobeyed for image taking. If M r and M are smaller than these numbers, inother words, the resolution in I polar is coarser than in the original image I ,one can either rst smooth I then perform the mapping, or alternatively, rstmap I to polar coordinates with proper resolutions followed by smoothing anddownscaling in r or . Which approach to take depends on the aspect ratio of I polar .

2.3 Relation to the Normal Fourier Transform in 2D2.3.1 Innite Transform

To nd the relation between the polar and the normal Fourier transforms, oneneeds to know the relation of their bases. The basis function for normal Fouriertransform represents a plane wave:

12

ei k r =1

2eikr cos( k ) . (43)

where k is the wave vector and ( k,k ) and ( r, ) are the polar coordinates of kand r respectively. The basis function is dened on the whole space and can beexpanded according to the Jacobi-Anger Identity [6] as

12

ei k r = 12

eikr cos( k )

=

m = im

12

J m (kr )eim ( k )

=

m =

im

2k e im k k,m (r, ) (44)

where k,m is dened in (34) and is known as cylindrical wave function. (44)means that a plane wave can be decomposed into cylindrical waves of exactlythe same wavenumber. Conversely,

k,m (r, ) = 2

0

(

i)m

2k eim k1

2 eikr cos( k ) dk . (45)

That is, plane waves of the same wavenumber, with their phases properly shiftedaccording to the direction of the wavevectors, can be superposed to get a cylin-drical wave. Alternatively, one says that the (normal) Fourier transform of k,mis

F (k,m )(k ,k ) = (k k )(i)m 2k e

im k . (46)

which is nonzero only on a circle of radius k.Suppose a function f (r, ) is dened on the whole space and its normal

Fourier transform is C k, k (C stands for C artesian. k and k are written as

12


14/26

(a) (b)

Figure 4: Illustration of the relation of the polar Fourier coefficients and thenormal Fourier coefficients C k, k . (a) When the space is innite, P k,m is theFourier coefficient of C k, k with k as the variable. (b) When the space is nite,P nm is the weighted sum of the Fourier coefficients on different circles. Withzero boundary condition, the weight function is proportional to (60).

subscripts for consistence of notations here although they take continuous val-ues), it can be expressed as

f (r, ) = 0 20 C k, k 12 eikr cos( k ) kdkdk . (47)Substituting (44) into (47),

f (r, ) = 0 m = im k 1 2 20 C k, k e im k dk k,m (r, )kdk . (48)One can recognize immediately that the expression inside the square bracketsis just the polar Fourier transform of f (r, ). If it is denoted as P k,m , one has

P k,m =im

k1

2 20 C k, k e im k dk . (49)The relation is very simple. Except for the factor ( im / k), P k,m is just theFourier coefficient of C k, k by considering k as variable. See Figure 4 (a) foran illustration.

(49) can be rewritten as

P k,m =im

k k1

2 0 20 (k k)e im k C k , k k dk dk (50)for convenience of later discussion.

2.3.2 Transform on Finite Regions

The relation between the polar Fourier transform and the normal Fourier trans-form is very simple when they are dened on the whole space. Strictly speaking,

13


15/26

it is ambiguous to talk about their relationship when dened on a nite region

as the basis functions are dened on regions of different shapes. For convenienceof discussion, we consider such a situation here: The normal Fourier transformis dened on a rectangle that is centered at the origin and encloses the diskwhere the polar Fourier transform is dened. Let the area of the rectangle beA.

We rst try to get the expansion of a plane wave in {nm }on the disk. Asshown in (44), a plane wave can be expanded in k,m , which in turn can beexpanded easily in {nm }on the disk with the help of (28):

ei k r =m

im J m (kr )eim ( k )

=

m

im

n

a

0Rnm ()J m (k)d Rnm (r )eim ( k )

=n,m

im 2 a0 Rnm ()J m (k)d e im k nm (r, ) . (51)for r a. The expression inside the square brackets is the coefficient of J m (kr )in Rnm (r ). It can be explicitly expressed by making use of (17). If knm areselected with the zero boundary condition,

a0 Rnm ()J m (k)d = ( 1)n 2knm J m (ka )k2 k2nm (52)and we have

eik

r

=n,m

(1)n im 2 k nmJ m (ka )

k2 k2nm e im k nm (r, ) . (53)

This equation holds for any k , including those appearing in the normal Fouriertransform dened on the rectangle, which we denote as k 0 .

A function f (r, ) dened on the disk can be extended to the rectangle bypadding 1 . Let the normal Fourier coefficients for the padded function be C k 0 .On the disk,

f (r, ) =k 0

C k 01

A ei k 0 r . (54)

f (r, ) can as well be expanded in {nm },

f (r, ) =nm

P nm nm (r, ) . (55)

With the expansion (53), it is easy to get that

P nm =k 0

p(k 0 ; n, m ) C k 0 (56)

p(k 0 ; n, m ) = ( 1)n im2 A knm

J m (k0a)k20 k2nm

e im k 0 (57)

1 It can be proved that the padding scheme does not affect the relations as given in (56)and (61).

14


16/26

0 0.5 1 1.5 2 2.5 3 3.50

0.5

1

1.5

2

2.5

k0

(a) (b)

Figure 5: (a) k 0 J m (k 0 a )k 20 k 2nm as a function of k0 and (b) the real part of p(k 0 ; n, m )as dened in (57) for n = 8, m = 5 and a = 32, where knm = 0 .994.

for the zero boundary condition. We write the main parts of P k,m and P nmfrom (50) and (56) for comparison:

P k,m dk (k k) e im k C k , k , (58)P nm

k 0

J m (k0a)k20 k2nme im k 0 C k 0 . (59)

When the space becomes nite, the integral over the wave vector is replaced bya summation and the sharp function of the wavenumber (k k) is replaced bya more spreading one (see Fig. 4 (b) for an illustration):

J m (k0a)k20 k2nm

. (60)

which has its maximum absolute value at k0 = knm . According to the asymp-totic behavior of Bessel functions, it arrives to its rst zeros approximately at

|k0 knm | = /a and will oscillatingly decrease on both sides. Fig. 5 (a) showsthe absolute value of (60) multiplied by k0 , which comes from the fact thatthe number of pixels at radius k is approximately proportional to k. Fig. 5(b) shows the real part of p(k 0 ; n, m ). One can compare it with the schematicillustration in Fig. 4 (b).

For completeness, if {nm }is determined with the derivative boundary con-dition, one hasP nm = ( 1)n im

2 A

aknm

k2nm a2 m2 k 0kJ m (ka )k2 k2nm

e im k C k 0 . (61)

(56) and (61) can be used to calculate the polar Fourier coefficients P nmfrom the normal Fourier coefficients C k 0 , which can be obtained by FFT. This

15


17/26

approach implies sinc interpolation in the spatial domain and is best suited

when the underlying original signal is band-limited.

2.4 Comparison with Zernike Polynomials

Basis functions dened with (35) are surely not the only existing orthogonalbasis. Actually as {Rnm |n = 1 , 2, }for any m forms a basis, one can randomlycombine Rnm with m () and still get an orthogonal basis for functions denedon the disk. But the choice of (35) is the most natural one. It has a clearphysical meaning, with the value of k indicating the scale. Apart from the Besselfunctions being radial functions, there exists, of course, also an innity of setsof basis functions on a disk. One of the most famous are Zernike polynomials.Since Teh et al. [11] made a comparison study on different moment methods,which shows that Zernike moments outperform other moment-based methodsin terms of overall performance, there are a lot of applications using Zernikemoments, e.g. [1315]. Zernike functions are dened on a unit disk, and, whenexpressed in polar coordinates, have the following form [3]

V nm (r, ) = Z nm (r )eim (62)

where m is any integer, n 0 is an integer and is the order of the polynomial,n |m|, n |m | is even. The angular part is the same as that of (35). Theradial Zernike function Z nm is a polynomial in r :

Z nm (r ) =

n | m |2

s =0(1)s

(n s)!s! n + |m |2

s ! n | m |2

s !

r n 2s . (63)

It has ( n |m|)/ 2 zeros between 0 and 1. The orthogonality relation of theradial functions is given by

10 Z nm (r )Z n m (r )rdr = 12n + 2 nn . (64)For purpose of comparison, we dene the normalized radial function as

Z nm = 2n + 2 Z nm . (65)The rst few normalized radial functions for m = 2 are shown in Fig. 1

(c). The typical form of Z nm with relatively large m and n is shown in Fig.

6 together with Rnm for comparison. From this gure, one can nd that Z nmhas also a wave-like form. Like Rnm , it has also a delay near the originfor the wave to begin; Unlike Rnm , the amplitude of the wave does notdecrease monotonically, instead the wavelength decreases with r . As {Rnm }and {Z nm }are both complete bases, they can be expressed by each other withthe help of the following relationship [3]:

10 Z nm (r )J m (xr )rdr = ( 1) n m2 J n +1 (x)x (66)where m 0 are integers.

16


18/26

0 0.2 0.4 0.6 0.8 13

2

1

0

1

2

3

4

r

(a)

0 0.2 0.4 0.6 0.8 14

2

0

2

4

6

8

r

(b)

Figure 6: (a) R (11)8 as dened in (27) with derivative boundary condition and(b) Z (28)8 as dened in (65). Both have 10 zeros on (0 , 1).

17


19/26

3 Spherical Fourier Tranform

3.1 Basis Functions and Expansion

We will follow the same approach as for polar coordinates, and much of thediscussion there also applies here. The expression of the Laplacian in sphericalcoordinates is given by

2 =

2r +

1r 2

2 (67)

where the radial part is

2r =

1r 2

r

r 2 r

(68)

and the angular part is

2 = 1sin

sin

+ 1sin2

22

. (69)

The Helmholtz equation is then given by

2r (r, , ) +

1r 2

2(r, , ) + k

2(r, , ) = 0 . (70)

For a solution of the form ( r, , ) = R(r )(,), one has

(,) = Y lm (,) (71)

where Y lm is a spherical harmonic as dened in (5). It satises

2Y lm + l(l + 1) Y lm = 0 . (72)

The corresponding radial part satises

1r 2

r

r 2 r

R + k2 l(l + 1)

r 2R = 0 . (73)

Its non-singular solution isR(r ) = j l (kr ) (74)

where j l is the so-called spherical Bessel function of order l and is related to theordinary Bessel Function by

j l (x) = 2x J l+

12 (x) . (75)

The spherical Bessel functions satisfy the orthogonality relation

0 j l (k1r ) j l (k2r )r 2dr = 2k21 (k1 k2) . (76)Putting the radial part (76) and angular part (71) together, one get the normal-ized basis function for spherical Fourier transform dened on the whole space

k,l,m (r, , ) = 2 kj l (kr )Y lm (,) (77)18


20/26

When the integral is only over a nite region [0 , a], the orthogonality relation

generally will not hold, instead

a0 j l (k1r ) j l (k2r )r 2dr = a2k21 k22 [k2 j l (k1a) j l (k2a) k1 j l (k2a) j l (k1a)] . (78)One needs boundary conditions to select a set of orthogonal basis functions. Forr = a the S-L boundary condition is

R(a)cos a2R (a)sin = 0 (79)with [0, 2). With R(r ) = j l (kr ), the above boundary condition becomes

j l (ka )cos (ka ) j l (ka )a sin = 0 . (80)Set x = ka and absorb the extra a into the choice of , the boundary conditionbecomes

j l (x)cos j l (x)sin = 0 . (81)If x l1 < x l2 < < x ln < are the nonnegative solutions to (81) with j l (x ln r/a ) nonzero, one can dene

knl =x lna

.

The n-th eigenfunction is then j l (knl r ). The orthogonal relation of the eigen-functions is

a

0 j l (knl r ) j l (kn l r )r 2dr = N ( l)n nn . (82)

It can be shown that

N ( l)n =a3

2 j l

2(x ln ) +1

x ln j l (x ln ) j l (x ln ) + 1

l(l + 1)x2ln

j 2l (x ln ) . (83)

With the zero-value boundary condition , (x l1 , x l2 , , x ln , ) are the pos-itive zeros of j l (x) andN ( l)n =

a3

2j 2l+1 (x ln ) . (84)

With the derivative boundary condition , (x l1 , x l2 , , x ln , ) are the pos-itive zeros of j l (x) except for x01 = 0. And

N ( l)n = a3

21 l(l + 1)x2ln j

2l (x ln ) (85)

with the special case

N (0)1 =a3

3. (86)

The normalized radial basis functions can be dened as

Rnl (r ) =1

N ( l)n j l (knl r ) . (87)19


21/26

Together with the angular part, the whole basis function for a solid sphere of

radius a can be dened asnlm (r, , ) = Rnl (r )Y lm (,) . (88)

A function f (r, , ) dened on a r a can be expanded in terms of nlm (r, , ):f (r, , ) =

n =1

l=0

l

m = l

S nlm nlm (r, , ) (89)

where

S nlm = a0 0 20 f (r, , )nlm (r, , )r 2 sin drdd (90)are the spherical Fourier coefficients ( S stands for S pherical). For real-valuedfunctions,

S nlm = S nl ( m ) . (91)

The discussion about mapping an image from Cartesian coordinates to polarcoordinates in last section also applies here. The only difference is that for asolid sphere of radius a (in the unit of a voxel size), the safe size in sphericalcoordinates should be 3a, 3a and 2 3a for r , and respectively.

3.2 Relation to the Normal Fourier Transform in 3D

A plane wave in 3D can be expanded in spherical waves k,l,m (r, , ) as [7]

1

23

eik

r

= 2

l=0

l

m = li l j l (kr )Y lm (,)Y lm (k ,k )

=1k

l=0

l

m = l

i l Y lm (k ,k )k,l,m (r, , ) (92)

where (k, k ,k ) are the spherical coordinates of the wave vector k .Any function f (r, , ) dened on the whole space can be expanded in either

of the two bases:

f (r, , ) = 0 0 20 C k, k , k 1 2 3 ei k r k sin k dk dk dk (93)=

l=0

l

m = l

0S k,l,m k,l,m (r, , ) k dk . (94)

With the relation of the bases as given in (92), one can easily get the relationof the coefficients:

S k,l,m =i l

k 0 20 C k, k , k Y lm (k ,k )sin k dk dk . (95)Except for a constant factor, S k,l,m is the SH coefficient of C k, k , k with (k ,k )as variables.

20


22/26

A function f (r, , ) that is dened on a solid sphere of nite radius a can

be expanded either in normal Fourier series or in spherical Fourier series. Herethe normal Fourier series is dened on a rectangular box which contains thesolid sphere and has its center also at the origin. Suppose the volume of therectangular box is V .

f (r, , ) =k 0

C k 01

V ei k r (96)

=nlm

S nlm nlm (r, , ) . (97)

The Fourier coefficients have the following relationship if knm are selected withzero boundary condition:

S nlm = (1)n i l 4 2aV k 0 knl j l (k0a)

k20 k2nmY lm (k 0 ,k 0 ) C k 0 . (98)

4 Applications of Polar and Spherical FourierTransforms

The polar and the spherical Fourier transforms can be regarded as variantionsof the Fourier transform. They can have applications in different problems. Asthe basis function is made up of the radial and the angular part separately, itis easy investigate how the transform coefficients change when the function isrotated.

If a 2D rotation operator R( ) is dened byR()f (r, ) = f (r, ), (99)

it works on a basis function of the polar Fourier transform as

R( )nm (r, ) = R( )Rnm (r )() (100)= Rnm (r )e im () (101)

= e im nm (r, ) . (102)

When it operates on a function f (r, ) with polar Fourier coefficients P nm , thechange of the function can be regarded as the change of the coefficients P nm .

R( )f (r, ) = R()

n =1

m = P nm nm (r, ) (103)

=

n =1

m = P nm R( )nm (r, ) (104)

=

n =1

m = P nm e im nm (r, ) (105)

With the rotation of R(), the coefficientP nm =P nm e

im .

21


23/26

The phase changes carry the information of rotation, therefore can be used to

estimate the rotation. This property can be employed for registration of images.Under rotation, only the phase of P nm is changed, its magnitude remainsthe same and is therefore a rotational invariant of the function. It will becalled a Polar Fourier Descriptor (PFD). The transform coefficients providea complete representation of the original function. Theoretically, a completeset of rotational invariant descriptors can be obtained by properly normalizingthe coefficients according to the degree of rotational symmetry (similar to thetechnique in [22]). However, although phase information is very important, therestill lacks a systematic and robust way of incorporating this information intothe descriptors. By discarding the phases, PFDs are no longer mathematicallycomplete. Nevertheless they still make up a robust set of rotation invariantdescriptors.

Rotation in 3D is more complicated than in 2D as there are two angularcoordinates now. It is well known that Y lm with m = l, l + 1 , , l span asubspace that is invariant with respect to the rotation group. When the operatorR(,, ) ( , and are the Euler angles that represent the rotation) applyto Y lm , one has

R(,, )Y lm (, ) =l

m = l

D ( l)m m (,, )Y lm (, ). (106)

where D ( l)m m (,, ) are the Wigner-D functions. Its exact expression, togetherwith the corresponding denition of the Euler angles, can be found in [4] andwill not be given here. As all the variance of the basis function nlm underrotation is captured by its angular part, one can simply replace Y lm with nlmin (106) and the equation still holds.

R(,, )nlm (r,, ) =l

m = l

D ( l)m m (,, )nlm (r,, ) . (107)

When a function f (r,, ) with spherical Fourier coefficients S nlm is under ro-tation R(,, ),

R(,, )f (r,, ) (108)= R(,, )

n =1

l=0

l

m = l

S nlm nlm (r,, ) (109)

=

n =1

l=0

l

m = l

S nlml

m = l

D ( l)m m (,, )nlm (r,, ) (110)

=

n =1

l=0

l

m = l

l

m = l

D ( l)mm (,, )S nlm nlm (r,, ) (111)

D ( l)m m (,, ) with m, m = l, l + 1 , , l form a (2l + 1) (2l + 1) matrix

D ( l) (,, ) = D ( l)m m (,, ) ,

22


24/26

called the Wigner-D matrix. Dene S (nl ) as the column vector made up of S nlm

with m = l, l + 1 , , l. With the rotation R(,, ),S (nl ) =D ( l) (,, )S (nl ) .The rotation parameters are coded into the change of the spherical Fouriercoefficients, and the latter, can be used to estimate the former in turn.

The rotation operator unitary, and its representative matrix D ( l) (,, ) isthen also unitary. It means the magnitude of the vector S (nl ) remains unchangedunder rotation. Therefore

S (nl ) = lm = l |S nlm |2 = lm = l S nlm S nlm (112)is a rotation-invariant property of the object. We call it a Spherical Fourier

Descriptor (SFD). A SFD is indexted by two numbers: n and l.

5 Conclusion and Outlook

We propose to use the eigenfunctions of the Laplacian that are separable inpolar and spherical coordinates as basis functions for Image analysis. This ideaputs the proposed polar and spherical Fourier transform and the normal Fouriertransform into the same framework and ensures close resemblance and relationbetween them.

The changes of the transform coefficients under rotation can be simply ex-pressed as functions of the rotation parameters. This property can be usedto estimate rotation angles, which is essential in image registration. We have

also shown how rotation-invariant descriptors can be dened on the transformcoefficients.We have discussed how to calculate the coefficients by mapping the data

from Cartesian coordinated into polar or spherical coordinates. The angulartransforms can be done efficiently with fast programs [23,24], The radial trans-form has to be calculated for every coefficient independently. This is surelynot an efficient way. Whether fast algorithms exist for radial transforms is aquestion to be answered.

References

[1] N. N. Lebedev, Special functions and their applications , (Translated from

Russian by R.A.Silverman) chapter 5, pp. 98-142, Dover, 1972[2] W. Kaplan, Advanced Calculus , pp. 696-698, Addison-Wesley, 1991

[3] M. Born and E. Wolf, Principles of Optics: Electromagnetic Theory of Propagation, Interference, and Diffraction of Light , 7th ed. pp. 523-525,Cambridge, 1989

[4] M.Tinkham, Group Theory and Quantum Mechanics , pp. 101-115, Dover,1992

[5] Wikipedia, Sturm-Liouville theory, http://en.wikipedia.org/wiki/Sturm-Liouville_theory

23


25/26

[6] Weisstein, Eric W, Jacobi-Anger Expansion. From MathWorld

A Wolfram Web Resource. http://mathworld.wolfram.com/Jacobi-AngerExpansion.html

[7] K.E. Schmidt, The expansion of a plane wave, http://fermi.la.asu.edu/PHY577/notes/plane.pdf

[8] S. Erturk, T.J. Dennis, 3D model representation using spherical harmon-ics, Electronics Letters , vol. 33, no.11, pp.951-952, 22 May 1997

[9] M. Kazhdan, T. Funkhouser and S. Rusinkiewicz, Rotation InvariantSpherical Harmonic Representation of 3D Shape Descriptors, Symposium on Geometry Processing , pp. 167-175, June 2003

[10] H. Huang, L. Shen, R. Zhang, F. Makedon, A. Saykin, and J. Pearlman, A

Novel Surface Registration Algorithm in Medical Modeling Applications,IEEE Trans. on Information Technology in Biomedicine , vol. 11, no. 4,pp. 474-482, 2007

[11] C.-H. Teh and R.T. Chin, On image analysis by the methods of moments,IEEE Trans. Pattern Analysis and Machine Intelligence , vol. 10, no. 4,pp. 496-513, 1988

[12] Y. Zana and R.M. Cesar-Jr.Face recognition based on polar frequencyfeatures, ACM Trans. Applied Perception , vol. 3, no. 1, pp. 62-82, 2006

[13] A. Khotanzad and Y. H. Hong, Invariant Image Recognition by ZernikeMoments, IEEE Trans. Pattern Analysis and Machine Intelligence , vol.12, no. 5, pp. 489-497, May 1990

[14] W.-Y. Kim and Y.-S. Kim, Robust Rotation Angle Estimator, IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 21, no.8, pp. 768-773, 1999

[15] E.M. Arvacheh and H.R. Tizhoosh, Pattern Analysis Using Zernike Mo-ments, Instrumentation and Measurement Technology Conference, 2005.IMTC 2005. Proceedings of the IEEE Volume 2, pp. 1574-1578, 16-19 May2005

[16] A. Averbuch, R.R. Coifman, D.L. Donoho, M. Elad, M. Israeli,Accurateand Fast Discrete Polar Fourier Transform, : Signals, Systems and Com-puters, 2003. Conference Record of the Thirty-Seventh Asilomar Confer-ence on, vol. 2, pp. 1933-1937, 2003

[17] A. Makadia, L. Sorgi and k. Daniilidis, Rotation estimation from sphericalimages, Rotation estimation from spherical images, Pattern Recognition,ICPR 2004, Proceedings of the 17th International Conference on , vol. 3,pp. 590-593, 23-26 Aug. 2004

[18] R. Skomski, J.P. Liu, and D.J. Sellmyer, Quasicoherent nucleation modein two-phase nanomagnets, Phys. Rev. B 60, pp. 7359-7365, 1999

[19] B. Pons, Ability of monocentric close-coupling expansions to describe ion-ization in atomic collisions Phys. Rev. A 63, pp. 012704, 2000

24


26/26

[20] R. Bisseling and R. Kosloff The fast Hankel transform as a tool in the solu-

tion of the time dependent Schringer equation, Journal of Computational Physics , vol. 59, no. 1, pp. 136-151, May 1985

[21] D. Lemoine, The discrete Bessel transform algorithm, J. Chem. Phys.101, pp. 3936-3944, 1994

[22] H. Burkhardt, Transformationen zur lageinvarianten Merkmalgewinnung ,Ersch. als Fortschrittbericht (Reihe 10, Nr. 7) der VDI-Zeitschriften, VDI-Verlag, 1979

[23] FFTW Home Page, http://www.fftw.org/

[24] Fast Spherical Harmonic Transforms, http://www.cs.dartmouth.edu/~geelong/sphere/

25

Fourier Analysis in Polar and Spherical.pdf

Documents