A Novel wavelet based threshold selection technique for Image ...

e tInternational Journal on Emerging Technologies 3(2): 52-55 (2012)ISSN No. (Print) : 0975-8364ISSN No. (Online) : 2249-3255A Novel wavelet based threshold selection technique for ImageDenoisingM. Zahid Alam and Dr. Ravi Shankar Mishra*Department of Electronics and Communication Engineering, AISECT University Bhopal, (MP)** Department of Electronics and Communication Engineering, SISTEC Bhopal, (MP)(Received 15 Augest, 2012 Accepted 18 September, 2012)ABSTRACT: This paper presents a wavelet threshold selection technique for noise filtering and comparisonof different noise reduction techniques used in image denoising. Additive noise is applied to the image anddifferent wavelet technique is used to remove the noise from the image. Finally all the different techniques arecompared for PSNR, SSIM improvement and the processing time required.Keywords: image denoising, wavelet transform, additive noise.I. INTRODUCTIONImage denoising has been a well studied problem inthe field of image processing. Yet researcherscontinue to focus attention on it to better the currentstate-of-the-art. Recently proposed methods takedifferent approaches to the problem and yet theirdenoising performances are comparable. A pertinentquestion then to ask is whether there is a theoreticallimit to denoising performance and more importantly,are we there yet? As camera manufacturers continueto pack increasing numbers of pixels per unit area, anincrease in noise sensitivity manifests itself in theform of a noisier image. We study the performancebounds for the image denoising problem. Our work inthis paper estimates a lower bound on the meansquared error of the denoised result and compares theperformance of current state-of-the-art denoisingmethods with this bound. We show that despite thephenomenal recent progress in the quality ofdenoising algorithms, some room for improvementstill remains for a wide class of general images, andat certain signal-to-noise levels. Therefore, imagedenoising has still possibilities.II. PREVIOUS WORKIn this section we present the recent works providesus some useful information for development of theproposed work.In this paper, Turgay Celik [1] proposes a noveltechnique for unsupervised change detection in multitemporal satellite images using principal componentanalysis (PCA) and k-means clustering. Thedifference image is partitioned into blocks. Orthonormal eigenvectors are extracted through PCA ofh× h non overlapping block set to create aneigenvector space Simulation results show that theproposed algorithm performs quite well oncombating both the zero-mean Gaussian noise andthe speckle noise, which is quite attractive for changedetectionin optical and SAR images.Sudha et al. [2] first presents a wavelet-basedthresholding scheme for noise suppression inultrasound images. Quantitative and qualitativecomparisons of the results obtained by theproposed method with the results achieved fromthe other speckle noise reduction techniquesdemonstrate its higher performance for specklereduction. T.Ratha et al.[3]they described andanalyzed an algorithm for cleaning speckle noise inultrasound medical images. MathematicalMorphological operations are used in this algorithm.This algorithm is based on Morphological ImageCleaning algorithm (MIC). The algorithm uses adifferent technique for reconstructing the featuresthat are lost while removing the noise. Formorphological operations it also uses arbitrarystructuring elements suitable for the ultrasoundimages which have speckle noise. Pierrick et al. [4]proposed a Bayesian Non Local Means-BasedSpeckle Filtering In their proposal, a new version ofthe Non Local (NL) Means filter adapted for USimages is proposed. Originally developed forGaussian noise removal, a Bayesian framework isused to adapt the NL means filter for speckle noise.Experiments were carried out on synthetic data sets

Alam and Mishra 53with different speckle simulations. Nonlocal Means-Based Speckle Filtering for Ultrasound Images ispresented by [5] In this method, an adaptation of thenonlocal (NL) means filter is proposed for specklereduction in ultrasound (US) images. Originallydeveloped for additive white Gaussian noise, wepropose to use a Bayesian framework to derive a NLmeansfilter adapted to a relevant ultrasound noisemodel. Quantitative results on synthetic data showthe performances of the proposed method comparedto well-established and state-of-the-art methods.Results on real images demonstrate that the proposedmethod is able to preserve accurately edges andstructural details of the image. M. I. H. Bhuiya et al.[6] presented Wavelet-Based Despeckling of MedicalUltrasound Images with The Symmetric NormalInverse Gaussian Prior In their proposal, an efficientwavelet-based method is proposed for despecklingmedical ultrasound images. A closed-form Bayesianwavelet based maximum a posteriori denoiser isdeveloped in a homomorphic framework, based onmodeling the wavelet coefficients of the logtransformof the reflectivity with a symmetric normalinverse Gaussian (SNI G) prior. A simple method ispresented for obtain-ing the parameters of the SNIGprior using local neighbors. Thus, the proposedmethod is spatially adaptive. Rajan et al. [7] In theirpaper they discuss the speckle reduction in imageswith the recently proposed Wavelet EmbeddedAnisotropic Diffusion (WEAD) and WaveletEmbedded Complex Diffusion (WECD). Both thesemethods are improvements over anisotropic andcomplex diffusion by adding wavelet based bayesshrink in its second stage. Both WEAD and WECDproduce excellent results when compared with theexisting speckle reduction filters.Normal Inverse Gaussian Prior In their proposal, anefficient wavelet-based method is proposed fordespeckling medical ultrasound images. A closedformBayesian wavelet based maximum a posterioridenoiser is developed in a homomorphic framework,based on modeling the wavelet coefficients of thelog-transform of the reflectivity with a symmetricnormal inverse Gaussian (SNIG) prior. A simplemethod is presented for obtain-ing the parameters ofthe SNIG prior using local neighbors. Thus, theproposed method is spatially adaptive. Rajan et al. [7]In their paper they discuss the speckle reductionin images with the recently proposed WaveletEmbedded Anisotropic Diffusion (WEAD) andWavelet Embedded Complex Diffusion (WECD).Both these methods are improvements overanisotropic and complex diffusion by addingwavelet based bayes shrink in its second stage. BothWEAD and WECD produce excellent results whencompared with the existing speckle reduction filters.III. IMAGE FILTRATION TECHNIQUES ANDTRANSFORMSA. Wavelet TransformA wavelet is a wave-like oscillation with amplitudethat starts out at zero, increases, and then decreasesback to zero. It can typically be visualized as a "briefoscillation" like one might see recorded by aseismograph or heart monitor. Generally, waveletsare purposefully crafted to have specific propertiesthat make them useful for signal processing.Fig. 1. Simple Wavelet.The wavelet transform replaces the Fouriertransform's sinusoidal waves by a family generatedby translations and dilations of a window called awavelet. It takes two arguments: time and scale. Adiscrete wavelet transform is any wavelet transformfor which the wavelets are discretely sampled.B. Wavelet FilteringAfter calculating the DWT coefficient form the DWTtransform the suppression or elimination ofcoefficients is performed there are many techniquesavailable for the selection of coefficients andsuppression of coefficients and then the inverse DWTis taken to generate the filtered image.C. Median FilterIn signal processing, it is often desirable to be able toperform some kind of noise reduction on an image orsignal. The median filter is a nonlinear digitalfiltering technique, often used to remove noise. Suchnoise reduction is a typical pre-processing step toimprove the results of later processing (for example,edge detection on an image). Median filtering is verywidely used in digital image processing because,under certain conditions, it preserves edges whileremoving noise (but see discussion below).The main idea of the median filter is to run throughthe signal entry by entry, replacing each entry with

Alam and Mishra 54the median of neighboring entries. The pattern ofneighbors is called the "window", which slides, entryby entry, over the entire signal. For 1D signals, themost obvious window is just the first few precedingand following entries, whereas for 2D (or higher -dimensional) signals such as images, more complexwindow patterns are possible (such as "box" or"cross" patterns). Note that if the window has an oddnumber of entries, then the median is simple todefine: it is just the middle value after all the entriesin the window are sorted numerically. For an evennumber of entries, there is more than one possiblemedian.IV. PROPOSED ALGORITHMThe proposed algorithm works on waveletthresholding and the threshold is selected byfollowing method.Let the sub band regions of the two-dimensional (2-D) critically sampled wavelet transform. Forconvenience, let us label the sub bands HH k , HL k ,and LH k, where k is the scale, and j is the coarsestscale. The smaller k is, the finer the scale is. Let usalso define sub-band P(S). P(S) is the sub-band of theparents of the coefficients of the sub-band S. Forexample, if S is HH 1 , then P(S) is HH 2 , or if S is HL 2 ,then P(S) is HL 3 .To estimate the noise variance σ 2 n from the noisywavelet coefficients, a robust median estimator isused from the finest scale wavelet coefficients (HH 1subband).2/ 0.6745 ‏(׀ y׀)‏ i σ n = MedianWhere y i is element of sub band HH 1 , σ y1 and σ y2can be found by :Using bivariate shrinkage functionEach coefficient is estimated.V. EXPERIMENTAL RESULTSThe algorithms are programmed in MATLAB and thesimulated results for PSNR(Peak signal to noiseratio) and SSIM(Structural Similarity) and processingtime, with different noise variance.Table 1: Comparison of various techniques (forimage distorted by noise of variance 0.05, thePSNR and SSIM of noisy image are 30.45dB and32.15% respectively).MethodPSNRTimeSSIM(%)(dB)(Sec.)LEE 32.51 62.81 5.17Median 31.22 61.35 0.147Adaptive 30.02 79.41 2.127Wavelet soft 27.96 41.23 1.18Wavelet hard 26.38 64.26 1.182Table 2: Comparison of various techniques (forimage distorted by noise of variance 0.01, thePSNR and SSIM of noisy image are 37.57dB and63.44% respectively).Where y1 and y2 are Variances of y 1 and y 2 .Using these variances signal variance 1 & 2 canbe estimated by applying formula given as:MethodPSNR(dB)SSIM(%) Time(Sec.)LEE 39.35 83.93 5.19Median 40.55 82.89 0.016Adaptive 41.23 90.18 0.021Waveletsoft33.86 73.35 0.42Wavelethard33.83 85.19 0.128

Alam and Mishra 55VI. CONCLUSIONThe simulation result shows that the wavelet domaintechniques take larger time than the spatialtechniques and also lags the PSNR and SSIMperformance (it adds the artifacts) at the lower noiselevels but at higher noise level the wavelet techniqueoutperforms them also the proposed methods willprovide good PSNR value.. In Principal ComponentAnalysis based Wavelet threshold technique, we willtry to enhance performance of Peak Signal to NoiseRatio by 3-4 dB and Structure Similarity (SSIM)around 70% with lesser Processing time.REFERENCES[1] Turgay Celik , “Unsupervised Change Detectionin Satellite Images Using Principal ComponentAnalysisand k-Means Clustering” IEEE Geoscienceand Remote Sensing Letters, Vol. 6, No. 4, October2009.[2] S. Sudha, G.R Suresh and R. Suknesh,"Speckle Noise Reduction in Ultrasound imagesBy Wavelet Thresholding Based On WeightedVariance", International Journal of ComputerTheory and Engineering, Vol. 1, No. 1, PP 7-12,2009[3] T. Ratha Jeyalakshmi and K. Ramar “A ModifiedMethod for Speckle Noise Removal in UltrasoundMedical Images” International Journal of Computerand Electrical Engineering, Vol. 2, No. 1, February,2010 1793-8163.[4] Pierrick Coupe, Pierre Hellier, Charles Kervrannand Christian Barillot “Bayesian Non Local Means-Based Speckle Filtering”, 1291-1294(ISBI 2008).[5] Pierrick Coupe, Pierre Hellier, Charles Kervrann,and Christian Barillot “Nonlocal Means-BasedSpeckle Filtering for Ultrasound Images” IEEETransactions On Image Processing, Vol. 18, No. 10,October 2009.[6] M.I.H. Bhuiyan, M. Omair Ahmad, Fellow, IEEE,and M.N.S. Swamy “Wavelet-Based Despeckling OfMedical Ultrasound Images With The SymmetricNormal Inverse Gaussian Prior” IEEE I, 721-724(ICASSP 2007).[7] Jeny Rajan and M.R. Kaimal “Speckle Reductionin Images with WEAD and WECD” ICVGIP 2006,LNCS 4338, pp. 184–193, 2006.