Medical Image Compression Using SFQ

8/2/2019 Medical Image Compression Using SFQ

1/27

4/22/2012

1

THESIS PROGRESS PRESENTATION

O N

Implementation of Efficient Medical Image

Compression Using Space-Frequency Quantization

Presented byKrishna Kumar

R.N.- 2010EL27

M. tech 4rd sem.

(DIGITAL SYSTEM)

Under The Guidance Of

Dr. Basant Kumar

Assist. Prof., ECED

MNNIT, Allahabad

Outline

(1) INTRODUCTION.

(2) EZW CODING STEPS.

(2) MDL SFQ CODER DESIGN STEPS.

(3) MODELLING OF SUB-BAND WAVELET

COEFFICIENTS.

(4) RESULTS .

( 5)CONCLUSION.


2/27

4/22/2012

2

Data Compression in TelemedicineData Compression in Telemedicine

Medical images require :

Larger memory space for storage

Higher bandwidth

Higher transmission time

Types of compression:

Lossless compression

limited compression ratio

Lossy compression

can be used with compromising quality.

Compression Technique Used

WAVELET TRANSFORM METHODS.

(TRANSFORM BASED IMAGE COMPRESSION).

ADVANTAGES :

Avoids blocking artefacts.

Facilitates progressive transmission of images.

Better matched to the hvs(human visual system)

characteristics

Compression scalable to achieve high compression ratios.

Very efficient at low bit rates.


3/27

4/22/2012

3

Step 1:Compute DWT Coefficients of the medical image.

Step 2:Set threshold and Iteration count.

Step 3:Apply Zero-tree pruning algorithm.

Step 4:Divide the survivor set of DWT coefficients into two sets of

positive and negative coefficients.

Step 5:Apply Lloyd-Max algorithm based uniform scalar quantizer

on positive and negative survivor sets.

Step 6:Apply Huffman code for encoding quantized DWT

coefficients.

Steps Followed in MDL SFQ Coder Design

Sub-band decomposition of an N x M image

H0

HH

H1

H0

H1

H0

HL

LH

LL

H1 2

2

2

2

2

2

a0

a1

N

MN/2

M/2

M/2

N


4/27

4/22/2012

4

Subband Structure (4- Level)

N/8

N/4

N/2

N

Scanning of wavelet coefficients for encoding

using EZW


5/27

4/22/2012

5

2. EZW encoding:Lets use 2-Level decomposition

26 6

-7 7

4

46

-4

-22

-34

0-2

1013

2log 260T = 2 = 16

sp zr zr zr

26sL

EZW encoding (Continued..)

The reconstructed value of this coefficient is

24 0

0 0

0

00

0

00

00

00

00

01.5 T = 24


6/27

4/22/2012

6


Correction of a two level quantizer with reconstruction level

T0/4.

28 0

0 0

0

00

0

00

00

00

00


Now w e reduce the th reshold by a factor o f 2and repeat th e pro cess.

* 6

-7 7

4

46

-4

-22

-34

0-2

1013

01

TT = = 8

2

iz zr zr sp sp iz iz


7/27

4/22/2012

7


Now th e t hreshold coeff icients are reconstr uctedw ith values 1.5T1=12.

28 0

0 0

0

00

0

00

00

00

1212

sL = 26,13,10


26 0

0 0

0

00

0

00

00

00

1014

0TCorrection becomes = 24


8/27

4/22/2012

8

Scalar Quantization

M ean square quant izat ion error

To design of q uant izat ion m eans deter m ine t heboundar ies bj and level yj tha t m in im izesquant izat ion error .

(1) i

i-1

bM 22

q x

j=1 b

= f (x)dxix-y

Scalar Quantization(Continued..)

These values of quantization can be

determined with the help of Lloyd-Max

quantizer.

It is a pdf optimized quantizer


9/27

4/22/2012

9

Scalar Quantization(Continued..)

The design equation for Lloyd-Max quantizer is

j

j-1

j

j-1

b

xb

j b

xb

x f x dx

y =f x dx

j+1 j

j

y + yb =

2

Needs For Statistical Modelling of Medical

Image

To characterized the image in the transform domain.

Benefits in model dependent quantization scheme

Characteristic of medical image wavelet coefficients:

Peaky

Heavy-tailed

Non-gaussian statistics


10/27

4/22/2012

10

Some heavy-tailed distributions

Generalized Student-t distribution

Generalized Pareto distribution

Weibull distribution

Gamma distribution

Gamma Distribution

: standard deviat io n


11/27

4/22/2012

11

Comparison of Various Distribution

Level 2 Distribution Chi-square value

(positive data)

Chi-square value

(negativedata)

HL 2 Generalized Pareto 0.9980 0.5692

Weibull 1.0687 0.5417

Generalized student-t 3.1374 0.2028

Gamma 0.9631 0.5231

LH 2 Generalized Pareto 1.2523 0.5977

Weibull 1.2552 0.5714


Gamma 1.0240 0.4985

HH 2 Generalized Pareto 0.8593 0.5741

Weibull 0.8792 0.5473


Gamma 0.8315 0.4996

Chi- Square Value (+Ve Data)

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

HL2 LH2 HH2

G.Pareto

W e ibu l l

G.Student- t

G amma

Sub-

bands


12/27

4/22/2012

12

Chi- Square Value (-Ve Data)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

HL2 LH2 HH2

G.Pareto

W e ibu l l

G.Student- t

G amma

Sub-bands

Jointly Optimization of SFQ Quantizer

M DL Criter ion is used

Best m odel is th at w hich g ives m in im umdescript ion length .

The code length for f ind ing the b in indices

W here: m- no. of quant izat ion levels- b in w i d t hCk- no. of coef f i c ients in Kth b insn- l engt h o f subband

2( , ) log kk

CL X m C

n


13/27

4/22/2012

13

Jointly Optimization of SFQ Quantizer(Continued..)

M DL cr iter ion can be w r i t ten as

2

2, 12

1( , ) ( ) ( , )

2 log

n

ij ij

i je

L X X x x L X m

Jointly Optimization of SFQ Quantizer

(Continued..)

To optimize jointly the two quantization we

use the principle of bit allocation,

The best performance is achieved when the

two quantizer operated on the same slope

(=R/D) on RD curve.


14/27

4/22/2012

14

RESULTSPerformance Comparison

IM AGES BPP=.25 BPP=.5 BPP=1

SPIHT M DL PROPOSEDSFQ M DL-SFQ



US-1 42.10 42 .92 4 4 . 8 2 46 .81 47.53 4 9 . 1 1 51 .05 51.40 5 2 . 9 0

US-2 43.27 45 .6 4 6 . 9 2 51 .7 52 .01 5 3 . 1 3 51 .90 52.40 5 4 . 1 1

CT-140.32 42 .47 4 3 . 2 3 43 .67 45.44 4 6 . 7 7 47 .82 49.00 5 0 . 4 2

CT-2 40.58 42 .24 4 2 . 9 6 42 .79 44.77 4 6 . 2 2 47 48 .35 4 9 . 6 7

MR-1 39.80 41 .27 4 2 . 0 7 43 .31 44.17 4 5 . 3 0 44 .98 46.27 4 7 . 5 3

MR-2 39.00 39 .81 4 0 . 0 1 41 .70 42.72 4 3 . 8 7 43 .30 44.82 4 6 . 4 2

PSNR Comparison for US-1 Image

34

36

38

40

42

44

46

48

BPP=0.25 BPP=0.5 BPP=1

SPIHT

M DL-SFQ

Proposed SFQ


15/27

4/22/2012

15

PSNR Comparison for US-2 Image

34

36

38

40

42

44

46

48


SPIHT

M DL-SFQ

Propo sed SFQ

PSNR Comparison for CT-1 Image

34

36

38

40

42

44

46

48


SPIHT

M DL-SFQ

Propo sed SFQ


16/27

4/22/2012

16

PSNR Comparison for CT-2 Image

34

36

38

40

42

44

46

48


SPIHT

M DL-SFQ

Proposed SFQ

PSNR Comparison for MR-1 Image

34

36

38

40

42

44

46

48


SPIHT

M DL-SFQ

Proposed SFQ


17/27

4/22/2012

17

PSNR Comparison for MR-2 Image

34

36

38

40

42

44

46

48


SPIHT

M DL-SFQ

Proposed SFQ

Image US-1


18/27

4/22/2012

18

Image US-1graph: psnr vs. bpp

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

50

bpp

PSNR

MDL SFQ

SFQ

Prosed MDL SFQ

IM AGE US-2


19/27

4/22/2012

19

IM AGE US-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

bpp

P

SNR

SPIHT

MDL SFQPropose MDL SFQ

IM AGE: CT-1


20/27

4/22/2012

20

IM AGE: CT-1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

50

bpp

PSNR

MDL SFQ

Propose MDL SFQ

IM AGE: CT-2


21/27

4/22/2012

21

IM AGE: CT-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

50

bpp

PSNR

MDL SFQ

Proposed MDL SFQ

IM AGE: M R-1


22/27

4/22/2012

22

IM AGE: M R-1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

50

bpp

PSNR

MDL SFQ

Propose MDL SFQ

IM AGE: M R-2


23/27

4/22/2012

23

IM AGE: M R-2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

5

10

15

20

25

30

35

40

45

bpp

PSNR

Proposed MDL SFQ

MDL SFQ

Conclusions

PROPOSED MDL-SFQ GIVES 1.18dB(APPROX)

IMPROVEMENT OVER EXITING MDL-SFQ.

PROPOSED MDL-SFQ GIVES IMPROVED RESULT ON

US,CT and MR IMAGES.

IMPROVEMENT IN PSNR OF CT AND MR IMAGES ARE

VERY LESS AS COMPARED TO US IMAGE


24/27

4/22/2012

24

Fur t her Wo rk

Im plem ent ation of ROI SFQ Coder.

REFERENCES

[ 1] .XlON G, Z., RAM CHAN DRAN , K., an d ORCHARD, M . T. (1997):' Sp ace f req u en cy q uan t izat io n f or w ave let im a ge co din g' , IEEETra ns.Im ag e Pro cess., 6, p. 67 7 6 93.

[ 2] .XlON G, Z., RAM CHAN DRAN , K., an d ORCHARD, M . T. (1998):'W avelet packet im age cod ing using space-f requency

quant izat ion ' ,IEEE Tra ns. Im ag e Process., 7, p p. 8 92 898

[ 3 ]. RA JPO OT, N . M . , W ILSO N , R. G., M EYER, F. G. , an d CO IFM A N , R.

R.,(2 00 3): ' Ad ap t iv e w a ve le t p acke t b asis se le ct i on f o r ze ro t r eeimage cod ing ' , IEEE Tra ns. Im ag e Process., 12, p p. 1 460 147 1


25/27

4/22/2012

25

REFEREN CES (CON T.)

[ 4] .PRZELA SKOW SKI, A., KAZU BEK, M ., an d JA M ROGIEW I CZ, T.

(1997):,Ef fect ive w avelet based com pression m et hod w it hadapt ive quant izat ion t hresho ld and zero -t ree cod ing ' .Proc.SPIE, M ult im edia Sto ra ge a nd Arch iving Syste m -II, 3229 , pp . 348

[ 5] .PRZELA SKOW SKI, A . (1 998 ):,' Fit t in g q uan t izat io n sch em e t om u l t ir eso lu t io n d et ail p re se rv in g co m p r essio n al go r it h m ' , Pr oc.

IEEE, p . 485 488[ 6] .PEA RLM A N W A . A SAID A , A n ew f ast an d ef ficie nt im a ge co de c

b ase d o n se t p ar t it i on in g in h ie rar ch ical t r ee s. ieee t ran sac t ions on circuits and system s for video t echno logy 199 6;6:243 50

[ 7] .SH ARIPO,M .9 (1 99 3): Em b e dd ed im a ge co d in g u sin g ze rot rees o f w avelet coef f icien t s. I EEE Tr ans. Signa l

Process.41,pp.3445-3462.

[8 ] .GERSHO,A.(1992):Pr inc ip les of quant izat ion. IEEE Trans. Oncircuit s an d syst em s.vol ca s-25,n o.-7.

[9] .ANTONINI,M . BARLAUD,M . M ATHIEU,P.DAUBECHIES,I.(1992): Im age cod ing using w avelet

transform. IEEE Tra ns. On im ag e p ro cessing .vol-1,no .-2



26/27

4/22/2012

26

[10] . L. Kau r ,R.C. Ch au han & S.C. Saxen a: Sp ace-f req ue ncyq u an t ise r d esign f or u lt r aso u nd im a ge co m p re ssi on b ase d o nm i n im u m d escr ip t io n le n gt h cr it e rio n . M edical & Biological Eng ineerin g & Com pu t ing 2005 , Vol. 43

[ 11 ].Vlad an Ve lisav lj ev ic, Balt asar Bef er ull-Lo zan o & M ar t in

Ve t te rli: Sace Fr eq u an ct Q uan t izat i on u sin g D ir ect i on le t s.ICIP IEEE 2007 .


BOOKS

[1].A WAVELET TOUR OF SIGNAL PROCESSING

(FIRST EDITION).

AUTHOR STEPHAN M ALLAT

[2 ].I NTRODU CTION TO DATA COM PRESSION

(THIRD EDITION).

AUTHO R KHALID SAYOOD


27/27

4/22/2012

27

Than k you

Medical Image Compression Using SFQ

Documents