Recent Development in Computational Intelligence and Engineering Applications, 16th Dec 2012, Guwahati, ISBN: 978-93-82208-49-5

38

AN ADAPTIVE STATIONARY WAVELET-BASED TECHNIQUE FOR DESPECKLING SAR IMAGE

AMLAN JYOTI DAS1, ANJAN KUMAR TALUKDAR2 & KANDARPA KUMAR SARMA3

1,2 & 3Dept. of Electronics & Communication Engineering, Gauhati University, Guwahati-14, Assam

Abstract- In this paper, we present a Stationary Wavelet Transform (SWT) based method for the purpose of despeckling the Synthetic Aperture radar (SAR) images by applying a maximum a posteriori probability (MAP) condition to estimate the noise free wavelet coefficients. A MAP Estimator is designed for this purpose which uses Rayleigh distribution for modeling the speckle noise and Laplacian distribution for modeling the statistics of the noise free wavelet coefficients. The parameters required for MAP estimator is determined by the technique used for parameter estimation after SWT. The experimental results show that the proposed despeckling algorithm efficiently removes speckle noise from the SAR images. Keywords- Synnthetic aperture radar (SAR), despeckling, Stationary Wavelet Transform (SWT), Maximum a posteriori probability (MAP) Estimator. I. INTRODUCTION Synthetic aperture radar (SAR) is a robust observation tool. It allows the acquisition of high resolution images of different places on the earth. SAR images are corrupted by special kind of noise known as speckle noise. In a SAR image, speckle manifests itself similar to thermal noise, by a random pixel-to-pixel variation with statistical properties. Therefore, in many SAR image processing operations like segmentation, speckle filtering is a crucial preprocessing step [1]. Many denoising algorithms have been developed for despeckling SAR images by using the Lee filter [2], the Frost filter [3], LG-MAP filter [13], the Gamma MAP filter [4], and their variations [5], [6]. These filters usually exhibit well in despeckling the SAR images, but they fail to provide sharp edge features and details of the original SAR image [7]. Since SAR images are multiplicative in nature, so many wavelet-based despeckling algorithms apply the log-transform to SAR images to statistically convert the multiplicative noise to additive noise prior to applying further denoising technique [7], [8]. An exponential operation is applied to convert the log-transformed images back to the nonlogarithmic format after wavelet denoising [8]. Several solutions have been proposed in the recent years, based on maximum a posteriori probability (MAP) criteria and different distributions: the gamma distribution [9], the α-stable distribution [10], the Pearson system of distributions [11], and the generalized Gaussian (GG) [12], laplacian and gaussian distribution [13] etc. In [12], a MAP criterion is derived which is associated with the Generalized Gaussian distribution and is performed in the undecimated wavelet domain. In [13], a MAP criterion is derived by considering Gaussian distribution for modeling speckle noise and

Laplacian distribution for modeling noise free wavelet coefficients. In [14], the noise-free image was approximated by a Gauss-Markov random field prior and the speckle noise was modeled using Gamma pdf. Although Discrete Wavelet Transform (DWT) plays a major role in the area of image denoising and image compression, the downsampling operation involved in DWT results in a time-variant translation and has to face difficulties in restoring original image discontinuities in the wavelet domain. Therefore, to restore the translation invariance property, lost by classical DWT, Stationary Wavelet Transform (SWT) has been preferred in many techniques [11]. In this paper we propose an efficient SWT based despeckling method by using MAP estimation. We avoid the log-transform and applied the MAP estimation criteria used as in [13], based on Gaussian distribution for modeling the speckle noise and Laplacian distribution for modeling the noise free wavelet coefficients. The parameter estimation is done using the results of [11]. The despeckling algorithms were tested on Ku-Band SAR image of pipeline over the Rio Grande river near Albuquerque, New Mexico (1-m resolution) with dimensions of 256×256, and were compared, which gives a satisfactory result for the proposed algorithm. This paper is organized as follows. In Section II, we describe statistical properties of SAR images, the two-dimensional (2-D) SWT algorithm, and signal modeling. In Section III, we have given a brief review of statistical characterization of wavelet coefficients and speckle noise. In Section IV we have shown the related work on MAP Estimator design. In Section V, the parameter estimation is described. Section VI provides a description of our algorithm and Section VII shows experimental results. Finally, a short conclusion is given.

Recent Dev

II. STATI

A. StatistiLet G be thand F be thmultiplicatexpressed image andrandom vawith unit m

B. StationThe SWT Figure1 salgorithm, lowpass filnoted in ththat the ouapplied to in Figure by two frolevel SWT doeMore predecimated convolvingwithout do

Figur

velopment in Com

ISTICAL MO

cal models for he observed sighe noise-free stive in nature,

as wd U is the normariable, follow

mean and varian

ary Wavelet Tralgorithm is sshows the

where anlters at level he figure that tutput of eathe input of th1, the two filtom filters of and . It is not includecisely, as DWDWT for a gi

g the signal wownsampling it

re 1: Block diagra

An Adaptive Stati

mputational Intellig

DELS FOR S

r SAR images gnal (intensitysignal. Since sp

the observed where, F is tmalized fadingwing a Gammnce . Its pd

Transform simple and is c2-D SWT

nd are the , respectively. the original imach decomposhe next levelters and the previous

is a redundante downsamplinWT, for leviven signal canwith the apprt.

am of SWT Deco

ionary Wavelet-Ba

gence and Enginee

SAR IMAGES

y or amplitude)peckle noise issignal can be

the noise freeg speckle-noisema distributiondf is given by

close to DWTdecompositionhighpass and

Also, it can bemage is andsition level is

. As shownare upsampleddecompositiont transform asng operationsel 1, all the

n be derived byropriate filters

mposition

ased Technique fo

ering Applications,

S

), s e e e n

T. n d e d s n d n s s. e y s

C. SignaThe obsfollowing

where free signaspeckle nspeckle c Let the signdecompoWe have,

Despecklmeans escoefficienwavelet timage. III. STAWAVELNOISE

The paramodelingand speckwe modcoefficienGaussiancompone D. GaussThe probdistribute

where, standard noise E. LaplaThe pdf odefined a

or Despeckling Sar

, 16th Dec 2012, G

al model served signal g model:

is the obseral that we wounoise, and component in th

be the stationanal x. It psition, where j,

ling an imagestimation of thnts and aptransform (ISW

ATISTICAL CLET COEFFI

ametric MAPg of prior probkle noise waveel the signal nts using the L

n model is uent.

sian model for bability densityed random vari

is the amplideviation of n

acian model forof a Laplacian as follows [13]

r Image

Guwahati, ISBN: 97

is assumed

rved signal,uld like to estim

denotes the she equivalent a

ary wavelet opperforms a j is the decom

in the multirehe speckle noipplying the inWT) to obtain

CHARACTERICIENTS AN

P estimator prbability distribelet coefficient

component Laplacian distr

used for mode

speckle compoy function (pdfiable , is defin

itude of the nnoise and is

r Signal Compdistributed ran

78-93-82208-49-5

to follow th

is the speckl

mate, is thsignal dependeadditive model

erator applied multiresolutio

mposition leve

solution domaise free wavel

nverse stationan the noise fr

RIZATION OND SPECKL

resumes propbution of signts. In this papeof the wavel

ribution, and theling the noi

onent f) for a Gaussianed as [15]:

noise, is ththe mean of th

onent ndom variable,

5

he

le-he

ent l.

to on el.

ain let

ary ree

OF LE

per nal er, let he ise

an

he he

,

Recent Dev

where determinesis the mean

IV. DESESTIMAT

Taking solution

The solutio

This solutibands (LHwavelet co V. PARAPROPOSE

velopment in Com

is the stands the spread ofn of the signal

SIGN OF TTOR

and equatin

on to this probl

ion is appliedH, HL, and HHoefficients.

AMETER ESED ESTIMAT

An Adaptive Stati

mputational Intellig

ard deviation f the density fu

.

THE PROPO

ng it to zero,

lem is given by

to the high fH) to estimate

STIMATIONTOR

ionary Wavelet-Ba

gence and Enginee

of signal, function and

OSED MAP

we have the

(14)y:

(15) frequency sub-the noise free

FOR THE

ased Technique fo

ering Applications,

P

e

-e

E

where,

where noisy wav

Inserting

where L is defi

2

2

(

( )

l kk

kk

h

h

ψ =

=

∑

∑ where adecompo (iii) coefficien

2lθσ ψ μ=

Using eqfor eachdecompo VI. EXP

The blockin Figurtransformimage thestimatedfrequencyeach decoused to cusing eqn The soluusing a p

or Despeckling Sar

, 16th Dec 2012, G

2Cθ =

is the normvelet coefficien

xC σ

μ=

(20) and (21) i2

2(

1F l

v

C ψσ =

+

is the look of tined as

2 2 2( 1)

2 2 2 1

) ( ) (Fo

) ( ) (Fo

lk

k

lk

k

g

g

−

−

∑

∑

and are the hsition level , r

(Variance onts): It can be c

22x Cθμ

qn. (16) - (22)h high freqsition level.

PERIMENTA

k diagram of tre 2. The or

med using SWThe parameters d using eqn. y subbands i.eomposition levcalculate the sn. (15) for each

ution of the Mprior knowled

r Image

Guwahati, ISBN: 97

is given b

2 2

2(1 )x l F

l F

C CC

ψψ

−

+ malized standant and is given

x

x

σμ into (19), we o

2

2 2 )x x

FCμ σ++

the SAR image

or diagonal subband

or horizontal and ve

highpass and lorespectively.

of the noisecalculated by u

the parametequency subb

AL DETAILS

the proposed mriginal noisy T. From the SW

mentioned in(16) - (22)

. LH, HL, andvel. These paraolution of the

h high frequenc

MAP estimatodge for model

78-93-82208-49-5

by

(18)ard deviation n by

(19) obtain

(20)

e, and paramet

d)

ertical subband)(2

owpass filters

e free wavelusing

(22)

rs are estimateands at eac

method is showimage is fir

WT decomposen Section V a) for the higd HH subband ameters are theMAP estimat

cy subbands.

r is derived bling the speck

5

of

ter

1)

at

let

ed ch

wn rst ed

are gh at en tor

by kle

Recent Dev

noise anddescribed iusing Gausis modeleconditions coefficientestimation the high frthe MAP high freqestimated wis taken festimated The resultshown in reported in VII. RESU For the ptechnique wgot by

Figur Applying thresholdinfiltering. methods inthresholds optimalitiespaces [1VisuShrinkbecause ituniversal tbe unwarranumber ocoefficientthreshold, coefficientFigure 3(bis worse thBayesShrin

velopment in Com

d noise free in Section III tssian distributi

ed using Lapare then

ts of LH, HLof the noise

requency subbestimator. The

quency subbanwavelet coeffi

for the low frhigh frequencs obtained by Figure 3. Som

n [17].

ULTS AND D

purpose of cowe have consid

re 2: Block diagra

existing tecng, BayesShrinAmong the

nclude VisuShrtake the advan

es over functi16]. Howevek yield overlyts threshold cthreshold and antedly large dof samples, wts. VisuShrinkwhich is globa

ts. The result ) which showshan other metnk threshold is

An Adaptive Stati

mputational Intellig

wavelet cothe speckle noion and the siglacian distribuapplied to

L, and HH sufree wavelet ands using theese results thends with thicients. Finallyfrequency subbcy subbands f

the proposedme of the rela

DISCUSSION

omparison of dered the resul

am of the propose

chniques likenk thresholdinmost popularrink and Bayentage of asympion spaces su

er, for imagy smoothed imchoice,

is the noise due to its depewhich remov

k provides a sally applied to of this metho

s that the qualitthods like Bays given by

ionary Wavelet-Ba

gence and Enginee

efficients. Asoise is modeledgnal componenution. Theseeach wavele

ubband for thecoefficients ofe conditions ofe formation of

he noise freey inverse SWTbands and thefor each leveld technique areated works are

the proposedts that we have

ed method.

e VisuShrinkng, and Wienerr thresholdingsShrink. Theseptotic minimaxuch as Besovge denoising

mages. This is (called the

variance), canendence on thees too manyingle value ofall the wavele

od is shown inty of the imageyesShrink. The

ased Technique fo

ering Applications,

s d

nt e t e f f f e T e l. e e

d e

k r g e x v g, s e n e y f t n e e

The propmethods comparinPSNR is

The PSNObepffafvariance a result a

Table 1

INPUT PSNR

27.6135

28.2838

29.7210

29.9965

or Despeckling Sar

, 16th Dec 2012, G

posed method like Bayes

ng the corresdefined as

NR is calculfenhofen withand applied tos mentioned in

Comparision oT

VISU SHRINK

BS

24.3189 2

24.1760 2

25.9526 2

26.0345 2

r Image

Guwahati, ISBN: 97

is compared wShrink threshponding PSN

ated for E-Sh different vao all the technin Table I.

of Psnr of Diffeechnique

BAYES SHRINK

WIENFILT

4.4568 24.4

4.7860 24.8

6.3420 26.8

6.7834 26.9

78-93-82208-49-5

with the existinholding etc b

NR values. Th

SAR images alues of noiques which giv

erent Denoisin

NER TER

PROPOSED

METHOD

756 28.3828

049 30.4351

564 30.9375

586 31.0219

5

ng by he

of

ise ve

g

S

D

8

1

5

9

An Adaptive Stationary Wavelet-Based Technique for Despeckling Sar Image

Recent Development in Computational Intelligence and Engineering Applications, 16th Dec 2012, Guwahati, ISBN: 978-93-82208-49-5

42

(a) (b)

(c) (d)

(e)

Figure 3: Original and denoised Ku-Band SAR image of pipeline over the Rio Grande river near Albuquerque, New Mexico: 1-m resolution. (a) Original noisy image (b) Denoised image using VisuShrink (c) Denoised image using BayesShrink (d) Denoised image

using Wiener filter (e) Denoised image using proposed method

An Adaptive Stationary Wavelet-Based Technique for Despeckling Sar Image

Recent Development in Computational Intelligence and Engineering Applications, 16th Dec 2012, Guwahati, ISBN: 978-93-82208-49-5

43

It is seen that the proposed method has better PSNR for all the input images with different values of noise variance. Thus the proposed approach is suitable for SAR applications. VIII. CONCLUSION

In this paper, an efficient technique for despeckling SAR images has been proposed. The speckle noise in the wavelet domain was modeled as an additive signal-dependent noise. We have tested the subband adaptive MAP processor, which relies on the Gaussian distribution of speckle noise and Laplacian prior for modeling the wavelet coefficients. Further, experimental results using the simulated speckled images show that, for five-level wavelet decompositions, the PSNR values of the proposed technique are higher than those algorithms mentioned above at less computational complexity. Although the speckle noise is removed by a significant amount, the denoised SAR image is smoothed in each successive decomposition levels. The subsequent stage of the proposed method will be related to sharpening the edges so that more edge information can be restored, enhancing the resolution of the denoised image.

REFERENCES [1] H. Xie, L. E. Pierce,and F. T. Ulaby “SAR speckle reduction

using wavelet denoising,” IEEE Transactions on Geoscience and Remote Sensing, Vol.40,No. 10, Oct. 2002.

[2] G. Lee, “Refined filtering of image noise using local statistics,” Comput. Graph. Image Process., vol. 15, no. 4, 1981.

[3] V. S. Frost, J. A. Stiles, K.S. Shanmugan, and J. C. Holtzman, “A model for radar images and its application to adaptive digital filtering of multiplicative noise,” IEEE Trans. Pattern Anal. Machine Intell.,vol.

PAMI-4, Mar. 1980.

[4] A. Lopes, E. Nezry, R. Touzi, and H. Laur, “Maximum a posteriori filtering and first order texture models in SAR images,” in Proc. IGARSS, 1990.

[5] C. Oliver and S. Quegan, “Understanding synthetic aperture radar Images.” Norwood, MA: Artech House, 1988.

[6] A. Lopes, R. Touzi, and E. Nezry, “Adaptive speckle filters and scene heterogeneity,” IEEE Trans. Geosci. Remote Sensing, vol. 28, pp. 992-1000, Nov. 1990.

[7] L. Gagnon and A. Jouan, “Speckle filtering of SAR images- A comparative study between complex-wavelet-based and standard filters,” Proc. SPIE, 1997.

[8] H. Guo, J. E. Odegard, M. Lang, R. A. Gopinath, I. W. Selesnick, and C. S. Burrus, “Wavelet based speckle reduction with application to SAR based ATD/R,” in Proc. ICIP, 1994.

[9] S. Solbø and T. Eltoft, “Γ-WMAP: A statistical speckle filter operating in the wavelet domain.” Int. J. Remote Sens., vol. 25, no. 5, pp. 1019-1036,Mar. 2004.

[10] A. Achim, P. Tsakalides, and A. Bezerianos, “SAR image denoising via Bayesian wavelet shrinkage based on heavy-tailed modeling,” IEEE Trans. Geosci. Remote Sens., vol. 41 , no. 8, pp. 1773-1784, Aug. 2003.

[11] S. Foucher, G. B. Bénié, and J.-M. Boucher, “Multiscale MAP filtering of SAR images,” IEEE Trans. Image Process., vol. 10, no. 1, pp. 49-60, Jan.2001.

[12] F. Argenti, T. Bianchi, and L. Alparone, “Multiresolution MAP despeckling of SAR images based on locally adaptive generalized Gaussian pdf modeling,” IEEE Trans. Image Process., vol. 15, no. 11, pp.3385-3399.Nov.2006.

[13] F. Argenti, T. Bianchi, A. Lapini and L. Alparone, “Fast MAP despeckling based on Laplacian-Gaussian modelling of wavelet coefficients,” IEEE Geoscience and Remote Sensing Letters, vol. 9, no. 1, Jan. 2012.

[14] M.Walessa and M. Datcu, “Model-based despeckling and information extraction from SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 38,no.5,pp.2258-2269,Sep.2000.

[15] Papoulis, A.: ‘Probability random variables and stochastic processes’ (MHL,NewYork,USA,1991).

[16] S. Grace Chang, Bin Yu, and Martin Vetterli, “Adaptive wavelet thresholding for image denoising and compression” IEEE Trans. On Image Processing, vol. 9, no. 9, Sept. 2000.

[17] A. J. Das, A. K. Talukdar, K. K. Sarma, “A novel adaptive stationary wavelet based technique for SAR image despeckling”, to appear in Proceedings of 1st International Conference on Mobile and Embedded Technology (MECON 2013), Noida, India, Jan, 2013.