Indonesian J our nal of Electrical Engineering and Computer Science V ol. 23, No. 3, September 2021, pp. 1440 1450 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v23.i3.pp1440-1450 r 1440 Image mixed gaussian and impulse noise elimination based on sparse r epr esentation model Ahmed Abdulqader Hussein 1 , Sabahaldin A. Hussain 2 , Ahmed Hameed Reja 3 1,3 Department of Electromechanical Engineering, Uni v ersity of T echnology , Baghdad, Iraq 2 Computer Science Department, Colle ge of Science, Aljufra Uni v ersity , Libya Article Inf o Article history: Recei v ed Dec 7, 2020 Re vised Jul 28, 2021 Accepted Aug 4, 2021 K eyw ords: Gaussian noise Image denoising Impulse noise Mix ed noise Sparse representation model ABSTRA CT A modified mix ed Gaussian plus impulse image denoising algorithm based on weighted encoding with image sparsity and nonlocal self-similarity priors re gulariza- tion is proposed in this paper . The encoding weights and the priors imposed on the im- ages are incorporated int o a v ariational frame w ork to treat more comple x mix ed noise distrib ution. Such noise is characterized by hea vy tails caused by impulse noise which needs to be eliminated through proper weighting of encoding residual. The outliers caused by the impulse nois e has a significant ef fect on the encoding weights. Hence a more accurate residual encoding error initialization pl ays the important role in o v erall denoising performance, especially at high im pulse nois e rates . In this paper , outliers free initialization image, and an easier to implement a parameter -free procedure for updating encoding weights ha v e been proposed. Experimental results demonstrate the capability of the proposed strate gy to reco v er images highly corrupted by mix ed Gaus- sian plus impulse noise as compared with the state of art denoising algorithm. The achie v ed results moti v ate us to implement the proposed algorithm in practice. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ahmed Abdulqader Hussein Department of Electromechanical Engineering Uni v ersity of T echnology , Baghdad, Iraq Email: ahmedabdulqaderhussein@gmail.com 1. INTR ODUCTION In v arious image applications, an image is ine vitably polluted by noise of dif ferent types. Mix ed noise could occur when an image that has already been contaminated by Gaussian noise in the procedure of image acquisition with f aulty equipment suf fers impulsi v e corruption during its transmission o v er noisy channels successi v ely [1]. The main objecti v e of an y denoising algorithm is to remo v e noise as totally as possible and to preserv e image edges as completely as possible. In a vie w of rele v ant studies, se v eral denoising algorithms ha v e been implemented for denoising a corrupted images by either impulse noise or Gaussian noise [2]-[6]. While, fe w rele v ant w orks ha v e been proposed in order to remo v e mix ed noise ef ficaciously due to the distinct characteristics of both kinds of de gradation processes. Generally , enhanced Gaussian denoising algorithms are considered to be insuf ficient for suppressing the impulse noise due to the noisy pix els are translated as an edges and should be preserv ed. Meanwhile, the strate gies for impulse noise elimination detect the impulse pix els and replace them with estimated v alues while lea ving the remaining pix els unchanged. Along these lines, most Gaussian noise in the restored images wil l be k ept leading to grain y and unsatisf actory visual results. Thus, a specific algorithm needs to be designed for standing ag ainst mix ed Gaussian plus impulse noise [7], [8]. The sparse representation (SR) has pro v ed its ef fecti v eness in dealing with v arious image processing J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1441 fields [9]-[11]. The algorithms base d on SR, benefit from the sparse nature of the images which can be well restricted by a linear combination of only fe w atoms of transformed domain called dictionary . The main weak- ness of SR model based denoising algorithms is its incapability in dealing with mix ed noise due to complicated properties of mix ed noise distrib ution. T o o v ercome SR model limitations, a dedicated algorithm need to be carefully designed to deal with hea vy tails of mix ed noise caused by impulse noise. Inte grating the SR into total v ariation frame w ork with proper encoding residual weighting, open the door for appearing algori thms capable of eli minating mix ed noise ef ficiently . A weighted encoding with sparse nonlocal re gularization model named (WESNR) w as successfully applied in [9]. The main goal of this paper is to enhance the performance of WESNR algorithm firstly by initial izing the algorithm with a more accurate image to ensure proper residual encoding error initialization and secondly by suggesting simpler weight updating procedure that ensures proper weighting of encoding residual to guarantee cancelling the hea vy tails of mix ed noise distrib ution induced by impulse noise. The achie v ed results sho w that the proposed algorithm e xhibits impro v ed performance in dealing with images highly corrupted by mix ed Gaussian plus impulse noise as well as f acilitating the practical implementation conditions. The rest of the paper is or g anized as follo ws. Related w ork is introduced in Section 2. Section 3 describes the mix ed Gaussian plus impul se noise model used in this paper . Section 4 briefly re vie ws the core of sparse denoising model. The proposed image denoising al gorithm is e xplained In Section 5. In section 6, the results of the proposed denoising algorithm is compa red with WESNR [9]. Finally , the concluding remarks are gi v en in section 7. 2. RELA TED W ORK A general adapti v e frame w ork for detection and elimination dif ferent types of noise, in v olving Gaus- sian noise, impulse noise and more importantly their mixtures has been proposed in [12]. The method is implemented based on maximum lik elihood estimation (MLE) approach as well as sparse representati on s o v er a trained dictionary . A pat ch-based approach for remo ving mix ed noise based on sparse representation has been presented in [13]. An optimization problem is formulated based on ` 1 -norm and ` 0 -quasiy-norm penalties. The sparse representation in a dictionary and sparsity of residual is enforced by ` 0 - ` 1 penalties respecti v ely . The dictionary is learned using independent component analysis (ICA). The image is denoised iterati v ely using a combination of soft and hard thresholding. Results sho w that the proposed method g a v e good results in terms of SSIM quantitati v e measure for images rich with fine details. Ho we v er , the performance of the proposed algorithm is not v erified under hea vily mix ed noise de gradation. A weighted encoding with sparse nonlocal re gularization (WESNR) that can deal with images corrupted by mix ed noise has been implemented in [9]. W eighted encoding is utilized in order to handle the impul se and Gaussian noise jointly . The prior image sparsity as well as prior nonlocal self-similarity ha v e been consolidated by means of a re gularization term and inserted into the v ariational encoding mechanism. The major dra wbacks of this algorithm is its incapability of dealing with highly corrupted images. A nonparametric Bayesian model to solv e sparse outlier has been proposed in [14], [15]. This strate gy pro v es its ef fecti v eness in treating mix ed Gaussian plus impulse (salt and pepper) noise. The noisy image is considered to be composed of three terms namely clean image, Gaussian noise image, and sparse impulse noise image. The model emplo ys spik e-slob sparse prior to recognize sparse coef ficients of the real data term and outlier noise. An algorithm for jointly denoising h yperspectral images corrupted by mix ed Gaussian plus impulse noise has ben proposed in [16]-[18]. The proposed algorithm a v ails the inherent spatial and spectral correlation of such images. The denoising problem is formulated based on synthesis prior (SP) and solv ed using Split-Bre gman algorithm. A ne w statistical re gularization term called joint adapti v e statistical prior (J ASP) has been prese n t ed in [19]. The proposed strate gy is incorporated into v ariational scheme. Moreo v er , the proposed v ariational model is solv ed using Split-Bre gman iterati v e algorithm to reco v er images corrupted by mix ed Gaussian plus impulse noise. A weighted joint sparse represe ntation model called WJSR to reco v er images corrupted by mix ed noise has been presented in [10]. The model group similar image patches and solv ed for global optimal solution using weighte d simultaneous orthogonal matching pursuit (W -SOMP). The weights are included to stands ag ainst impulse noise which abolishes image patch similarities. The WJSR is inte grated, with the global and sparse error priors, into v ariational frame w ork to ef ficiently deal with mix ed noise corruption. The major dra wback of this algorithm is its computation comple xity due to the usage of greedy algorithm W -SOMP . Ima g e mixed gaussian and impulse noise elimination based on spar se ... (Ahmed Abdulqader Hussein) Evaluation Warning : The document was created with Spire.PDF for Python.
1442 r ISSN: 2502-4752 An iterati v e non-con v e x lo w rank matrix approximation (NonLRMA) model for denoising h yper - spectral image (HIS) corrupted by mix ed noise has been proposed in [20]-[22]. NonLRMA decomposes the de graded HSI image into a lo w rank component and a sparse term. Results pro v e the capability of the proposed algorithm to preserv e lar ge-scale image structures and small-scale details o v er a wide range of image bands de gradations. An adapti v e Median based Non-local Lo w Rank Approximation (AMNLRA) approach for de- noising images polluted by mix ed g aussian plus impulse noise has been proposed in [23]. The proposed method is based on eliminating the ef fect of impulsi v e noise on the mix ed noise probability distrib ution to reco v er the Gaussian probabili ty distrib ution. T o locate the position of pix els contaminated by impulse noise, an adapti v e median filter AMF [24] w as used. According to data redundanc y , these identified pix els are then processed using non-local mean filtering (NLM) to con v ert the residual noise into a Gaussian noise distrib ution. Finally , the pre-processed image has been processed using reduced rank optimization to obtain the denoised image. 3. PR OBLEM FORMULA TION In a vie w of image denoising process, the elimination of mix ed noise is more complicated than the standardized si n gl e noise. As a rule, this noise is sampled from v arious distrib utions. In this paper , the proposed strate gy gi v es more attention for image denoising of mix ed noise which specified by a combination of Gaussian and impulse noise. Ob viously , it is a considerable problem that can be modeled as (1): y = I M P ( x + n ) (1) where, x , y and n denotes the original image pix el, corrupted image pix el and additi v e white Gaussian noise respecti v ely . The symbol I M P , denotes the corruption process by salt and pepper impulse noise. 4. SP ARSE DENOISING MODEL Let us denote an image by x 2 R N . The e xtracted image patch, can be represented as (2): x j = R j x (2) where x j the stretched v ector of an image is a patch of size p n p n and R j is the matrix e xtraction op- erator at location j . The goal of sparse representati on theory is to find an o v er complete dictionary D = [ d 1 ; d 2 ; ::::::: :d n ] 2 R n m to sparsely code x j where d j 2 R n is the j th atom of D . Based on the deri v ed dictionary D , the image patch can be re written according to (3) [9]: x j = D j (3) where j is a sparse coding v ector . The least square solution of x can be e xpressed as (4): x = D (4) where is the set of all coding v ectors j . In case of A WGN corrupted image y , can be e xpressed as (5) [9]: ^ = ar g min k y D k 2 2 + R ( ) (5) where R ( ) and are the re gularization term and re gularization parameter that ensure maximum a posteriori probability (MAP) solution of for A WGN model. Ho we v er , for mix ed noise remo v al, MAP solution is unreachable due to hea vy tail caused by impulse noise. T o deal with mix ed noise remo v al, a ne w model named weighted encoding with sparse nonlocal re gularization (WESNR) has been proposed. According to WESNR, The (5) is re written as [9]: ^ = ar g min W 1 2 ( y D ) 2 2 + R ( ) (6) where W is a diagonal weights matrix its elements W j j is deri v ed from residual error e j = ( y D ) j as [9]: W j j = exp ( c e 2 j ) (7) Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1440 1450 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1443 where c is a po s iti v e constant to control the decreasing rate of W j j . The weights W are adapti v ely updated in each iteration. The pix el corrupted by impulse noise is assigned small weight close to zero, while the uncorrupted pix el is left unchanged through assigning weight close to 1. This process will eliminate the ef fect of hea vy tails caused by impulse noise and hence ensure MAP solution of (6) to reco v e r images corrupted by mix ed Gaussian plus impulse noise. The updating of W depends on the coding residual error e . In [9], the initialized image x (0) for salt and pepper impulse noise is obtained by applying AMF [24] to y . Thus, the residual error is initialized according to (8). e (0) = y x (0) (8) Inspired by the w ork in [25], local sparsity and nonlocal self-similarity are the tw o priors adopted in WESNR to model re gularization term R ( ) yields to (9): ^ = ar g min W 1 2 ( y D ) 2 2 + k k 1 (9) as proposed in [9], the (9) w as solv ed via iterati v ely re weighted least square minimization scheme which has been proposed in [26] using the (10): v j j ( k + 1) = p ( j ( k ) j ) 2 + " ( k ) 2 (10) where j ( k ) and " ( k ) is the j th element of coding v ector and numerical stability parameter in the k th iteration respecti v ely . Hence, " ( k ) as well as are updated according to (11) and (12) respecti v ely [9]. " ( k + 1) = min ( " ( k ) ; median ( j ( k ) j )) (11) ( k + 1) = ( D T W D + V ( k + 1)) 1 ( D T W y D T W D ) + (12) 5. THE PR OPOSED ALGORITHM In WESNR [9], the weights W ha v e the important role in attenuating the ef fect of hea vy tails triggered by impulse noise and hence retrie v e back the distrib ution of mix ed Gaussian plus impulse noise to resemble Gaussian noise distrib ution. This will pa v e the w ay for sparse modeling technique to implement denoising algorithms that can deal with images corrupted by mix ed noise without e xplicit impulse detection phase. Re- ferring to (6), the accurac y of estimated sparse v ector ^ and t he diagonal weight matrix W are highly af fected by residual encoding error which in turn depends upon the ini tialized image x (0) . Thus, better performance is e xpected when more accurate image is used for initialization. Generally , the salt and pepper impulse noise tak e either minimum or maximum v alues in the dynamic range of the image (e.g. [0,255]). Thus, we need to process the pix els that tak e these e xtreme v alues. The question raised no w is ho w to replace the corrupted pix el in such a w ay that ensures resemblance with its noise free neighbor pix els and hence dimensions the ef fect of outliers. Intuiti v ely , critical replacements occurs when the test windo w u;z centered at pix el p u;z filled with impulses. T o solv e this problem, we suggest to adopt the Rob ust Outlier Exclusion (R OE) proposed in [27]. In this w ork the test windo w size is e xpanded to the ne xt higher one (typically 4 4) and apply the trimmed mea n which is characterized by its rob ustness ag ainst out liers. The conception for limiting the maximum test windo w size to 4 4 is to boost the estimated accurac y without influencing the significant correlation characteristics of neighbor pix els. As stated in [27], 25 % of the outliers e xcl usion accomplishes a po werful estimatation for reco v ering the corrupted pix els. Algorithm 1 sho ws the initialization algorithm steps according to R OE [27]. Ima g e mixed gaussian and impulse noise elimination based on spar se ... (Ahmed Abdulqader Hussein) Evaluation Warning : The document was created with Spire.PDF for Python.
1444 r ISSN: 2502-4752 Algorithm 1 Initialization algorithm Input: y , an image corrupted by impulse noise u;z Slide windo w centered at p u;z of Size M M Where M is an Odd number typically 3 Output: x (0) , the initialized image f or each non border pix el p ( i; j ) 2 y do if 0 < p u;z < 255 then The pix el is uncorrupted. Hence, left unchanged else The pix el is corrupted, do the follo wing: Exclude all noisy pix els in u;z Compute median v alue p median Set p u;z = p median if u;z is filled of impulse noise then Replace u;z by u;z of size M+1 M+1 Exclude all noisy pix els in u;z Compute T rim Mean in u;z with 25 % e xclusion p tr im Set p u;z = p tr im end if if u;z is filled of impulse noise then Compute the Mean in u;z p mean Set p u;z = p mean end if end if end f or Output the initialized image x (0) The initialization image x (0) is no w used to calculate the residual encoding error which plays i mpor - tant role in updating the weight matrix W . Referring to (7), the assigned weights W is firstly sensiti v e to the v alue of the constant c which controlli ng the decreasing rate of the weights and secondly use trigonometric e xponential function. T o modify the algorithm to suit the implementation requirements, it is usually best to a v oid the use of mathematical functions, in this case, e xponential functions. Hence, our goal is to replace the e xponential function while retaining the properties of its e x ecution. Referring to (7), impulse noise will cause a significant increase in residual error , which will lead to the assignation of weight close to zero ( W ! 0 ), thereby eliminating the impact of hea vy tails caused by impulse noise. Otherwise, if t he pix el has no impulse noise, the weight will be close to one ( W ! 1 ). No w to achie v e our goal, assuming a is a positi v e v alue greater than zero and less or equal to 1, then (1 a ) 2 ! 0 when a ! 1 and (1 a ) 2 ! 1 when a ! 0 . The function (1 a ) 2 has the same properties as the e xponential function. Consequently , the diagonal weight matrix W can be updated according to (13). W = (1 a ) 2 0 6 a 6 1 (13) The constraints imposed on the v ariable a in (13) can be r elated to the absolute normalized residual coding error ( j e nor m j ) . Accordingly , the weight matrix W is updated through the follo wing steps: 1. Calculate absolute residual error: e ( k ) = j ( y D ( k )) j (14) 2. Normalize the error by element wise di vision: e nor m ( k ) = e ( k ) max ( e ( k )) (15) 3. Update weight diagonal matrix: W j j ( k ) = (1 e nor m ( k ) ) 2 (16) Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1440 1450 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1445 Re g arding to the case of impulsi v e noise which is specified by (16), the normalized error ( j e nor m j ) approaches 1 and thus a v alue close to zero will be assigned to the encoding weights W j j which eliminate the ef fect of the impulsi v e noise. Hence, ensuring MAP solution of (6) in order to reco v er images corrupted by mix ed Gaussian plus impulse noise. Moreo v er , compared with (7) and (16) has the adv antage of being easier to implement in practice by a v oiding the use of an e xponential function. Therefore, there is no need to calibrate the parameter c . By follo wing the steps outlined abo v e, Algorithm 2 demonstrates the modification of the proposed WESNR based mix ed Gaussian plus impulse image denoising algorithm. Algorithm 2 Proposed mix ed noise denoising algorithm Input: Dictionary D , noisy image y , and number of iterations L Initialization : Apply R OE [27] to y in order to obtain the initialized image x (0) Calculate residual error e ( k = 0) using (14) Initialize to zero Output : Denoised image x f or k=1,2,....... L do 1- Compute ( k ) by (12) 2- Compute x ( k ) = D ( k ) 3- Update the nonlocal coding v ector = D T x ( k ) 4- Compute the residual error using (14) 5- Normalize the residual error using (15) 6- Update weight matrix W using (16) end f or Output the denoised image x = D ( L ) 6. EXPERIMENT AL RESUL TS The proposed algorithm is e v aluated on a set of standard images ha ving distinctly dif ferent features of size 512×512. These images are artificially polluted firstly by Gaussian noise with standard de viation of 10, 20, and 30 and then 30 % , 50 % , and 70 % of image pix els are replaced by sal t and pepper impulsi v e noise. Objecti v ely , the obtained res ults are e v aluated using peak signal to noise ratio (PSNR) and the structural similarity inde x (SSIM) [28]. Moreo v er , subjecti v e e v aluation is included to pro v e the quality of the reco v ered images. The proposed algorithm is compared with WESNR [9]. The parameters of WESNR are set to its def ault v alues as proposed in [9]. The parameters are c = 0.0008, (0) = 0.0001 to reduce the ef fect of impulse noise on block-matching then = 1 for > 10 or = 0.5 for 6 10 , " (0) = 0 : 1 and updated in accordance with (11). The image patches are encoded o v er a set of of fline learned PCA based dictionaries. The proposed algorithm has the same parameters as WESNR e xcept of the parameter c which has been discarded. The number of iterations L is k ept fix ed at 8 for all e xperiments. 6.1. Objecti v e e v aluation The denoising relati v e performance in terms of PSNR and SSIM for an a v erage of v e runs using a set of images are recorded in T able 2 (the v alue in parenthesis is the SSIM). According to the performance criteria, the superior outcomes are referred with a bold font. Re g arding to the results e xamination of T able 1, the proposed algorithm outperforms WESNR algorithm most of the time in terms of indi vidual, a v erage PSNR as well as SSIM v alues o v er the whole scope of noise le v els and images under test. In addition, by carefully checking the obtained results in T able 2, it is noticed that as the le v el of mix ed noise increases , the proposed algorithm e xhibits higher performance compared to WESNR. As an illustration, Figure 1a, Figure 1b, Figure 2a, and Figure 2b demonstrate the relati v e denoising of both algorithms in terms of PSNR and SSIM for =30, sp =30 % , and =30, sp =70 % respecti v ely . The achie v ed results pro v e the enhanced performance and the ef fecti v eness of the modification inserted on the con v entional WESNR algorithm namely starting with a more accurate initialized image x (0) . The more accurate initializat ion of the proposed algorithm has been reflected in increasing the accurac y of estimating the initialized weight matrix W . Furthermore, the proposed function for updating the diagonal weight matrix W in (16) pro v es its ability to attenuate the ef fect of hea vy tails caused by impulse noise considerably . Therefore, a MAP solution for reco v ering images corrupted by mix ed Gaussian plus impulse noise has been realized. Ima g e mixed gaussian and impulse noise elimination based on spar se ... (Ahmed Abdulqader Hussein) Evaluation Warning : The document was created with Spire.PDF for Python.
1446 r ISSN: 2502-4752 T able 1. PSNR (SSIM) denoising results Images sp 30 % sp 50 % sp 70 % WESNR Pr oposed WESNR Pr oposed WESNR Pr oposed Lena 10 33.31 (0.879) 33.84 (0.881) 31.79 (0.863) 32.50 (0.867) 29.31 (0.830) 29.85 (0.835) 20 30.68 (0.810) 31.03 0.818) 29.61 (0.796) 30.19 (0.807) 27.59 (0.769) 28.17 (0.787) 30 27.23 (0.634) 27.97 (0.657) 26.00 (0.603) 27.94 (0.633) 24.01 (0.577) 26.69 (0.620) Boat 10 30.56 (0.831) 31.39 (0.845) 28.73 (0.799) 29.56 (0.808) 26.12 (0.736) 26.87 (0.739) 20 28.27 (0.747) 28.68 (0.751) 26.89 (0.719) 27.54 (0.721) 24.86 (0.657) 25.49 (0.667) 30 25.83 (0.619) 26.39 (0.642) 24.34 (0.582) 25.93 (0.635) 22.29 (0.532) 24.45 (0.612) Hill 10 30.88 (0.815) 31.45 (0.830) 29.77 (0.785) 30.11 (0.791) 27.98 (0.724) 28.13 0.728) 20 28.74 (0.725) 29.01 (0.731) 27.86 (0.699) 28.11 (0.704) 26.39 (0.652) 26.57 (0.654) 30 26.27 (0.599) 26.81 (0.622) 25.22 (0.558) 26.54 (0.618) 23.60 (0.516) 25.62 (0.604) F16 10 32.47 (0.893) 29.46 (0.889) 30.61 (0.878) 28.84 (0.878) 27.72 (0.846) 27.20 (0.851) 20 30.14 (0.824) 28.18 (0.816) 28.72 (0.813) 27.68 (0.819) 26.32 (0.787) 26.21 (0.806) 30 26.79 (0.627) 26.32 (0.653) 25.65 (0.607) 26.20 (0.635) 23.28 (0.564) 25.22 (0.602) Monarch 10 35.36 (0.935) 35.78 (0.937) 33.09 (0.930) 34.29 (0.933) 28.72 (0.900) 29.83 (0.913) 20 31.66 (0.860) 32.04 (0.880) 30.03 (0.861) 31.16 (0.870) 26.68 (0.830) 28.10 (0.849) 30 27.51 (0.674) 28.31 (0.804) 25.92 (0.644) 28.24 (0.787) 23.00 (0.600) 26.15 (0.754) Barbara 10 29.18 (0.874) 31.44 (0.899) 26.99 (0.831) 28.82 (0.860) 23.97 (0.728) 25.21 (0.759) 20 27.13 (0.793) 28.65 (0.822) 25.44 (0.744) 26.65 (0.773) 22.75 (0.651) 23.88 (0.677) 30 24.57 (0.644) 25.96 (0.692) 22.64 (0.587) 25.02 (0.681) 20.49 (0.505) 23.20 (0.624) House 10 36.97 (0.912) 36.87 (0.902) 36.02 (0.904) 36.21 (0.900) 33.51 (0.886) 33.62 (0.890) 20 33.67 (0.843) 33.76 (0.862) 32.78 (0.837) 33.48 (0.859) 30.84 (0.825) 31.45 (0.848) 30 28.90 (0.630) 29.70 (0.777) 27.83 (0.610) 29.04 (0.713) 25.83 (0.598) 28.31 (0.654) Couple 10 30.34 (0.846) 31.17 (0.860) 28.77 (0.813) 29.50 (0.824) 26.18 (0.738) 26.77 (0.744) 20 28.04 (0.757) 28.52 (0.765) 26.84 (0.725) 27.32 (0.728) 24.77 (0.641) 25.20 (0.654) 30 25.56 (0.626) 26.12 (0.643) 24.37 (0.592) 25.63 (0.637) 22.51 (0.536) 24.19 (0.596) Man 10 30.71 (0.841) 31.18 (0.849) 29.36 (0.814) 29.80 (0.818) 27.18 (0.758) 27.61 (0.761) 20 28.51 (0.753) 28.74 (0.755) 27.65 (0.734) 27.91 (0.737) 25.69 (0.679) 26.14 (0.690) 30 26.09 (0.613) 26.60 (0.653) 24.86 (0.577) 26.25 (0.644) 22.97 (0.529) 25.23 (0.636) Zelda 10 34.73 (0.884) 34.99 (0.886) 33.83 (0.872) 34.06 (0.875) 32.10 (0.847) 32.23 (0.855) 20 31.87 (0.815) 32.14 (0.818) 31.05 (0.802) 31.53 (0.815) 29.52 (0.779) 30.09 (0.807) 30 28.10 (0.645) 29.76 (0.747) 27.14 (0.630) 28.48 (0.717) 25.60 (0.613) 27.71 (0.705) A v erage 10 32.45 (0.871) 32.76 (0.878) 30.90 (0.849) 31.37 (0.855) 28.28 (0.800) 28.73 (0.808) 20 29.87 (0.793) 30.08 (0.802) 28.69 (0.773) 29.16 (0.783) 26.54 (0.727) 27.13 (0.744) 30 26.69 (0.631) 27.39 (0.690) 25.40 (0.600) 26.93 (0.670) 23.36 (0.557) 25.68 (0.641) Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1440 1450 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1447 In order to complete the picture as a whole, Figures 3 and Figures 4 sho w the relati v e a v erage P SNR and SSIM performance for all images under the influence of dif ferent mixing noise le v els. In vie w of the accomplished results of Figures 3 and 4. the proposed algorithm performance has been dif ferentiated at = 30 compared to the WESNR algorithm. Along these lines, more accurate processing of impulse noise through R OE [27] has the greatest impact on the quality of denoised images, especially when sp e xceeds 50 % . (a) (b) Figure 1. Relati v e denoising algorithms performance for =30, sp =30 % ; (a) PSNR (dB) and (b) SSIM (a) (b) Figure 2. Relati v e denoising algorithms performance for =30, sp =70 % ; (a) PSNR (dB) and (b) SSIM (a) (b) Figure 3. Relati v e a v erage denoising performance for dif ferent mix ed noise le v els; (a) PSNR (dB) and (b) SSIM Finally , in order to quantitati v ely summarize the results in the case of high mix ed noise le v els of =30, sp =70 % , the proposed algorithm achie v es an a v erage PSNR g ain of approximately 2.32 dB o v er the entire set of images under se v ere noise conditions. Ima g e mixed gaussian and impulse noise elimination based on spar se ... (Ahmed Abdulqader Hussein) Evaluation Warning : The document was created with Spire.PDF for Python.
1448 r ISSN: 2502-4752 (a) (b) Figure 4. Relati v e a v erage denoising algorithms performance for =30; (a) PSNR (dB) and (b) SSIM 6.2. Subjecti v e e v aluation In accordance to the performance analysis of the proposed approach, it is significant to inte grate the visual comprehension in the relati v e performance comparison. Figure 5 illustrates a sample of images under test synthetically corrupted by dif ferent le v els of mix ed Gaussian plus impulse noise. The first column of Figure 5 sho ws the original image of Monarch, Barbara, and house respecti v ely . The subsequent columns demonstrate the noisy and the denoi sed images respecti v ely . Clearly , for =10 and sp =30 % , both algorithms achie v e nearly equi v alent performance (see Figure 5(a1-a4). The performance of the WESNR is de graded as the mix ed noise le v el increased. The de gradations become more ob vious when e xceeds 20 and sp e xceeds 50 % . At a such high mix ed noise en vironment, the proposed algorithm produces visually more pleasant results which can be easily noticed through Figure 5(b1-b4) and Figure 5(c1-c4). The proposed algorithm preserv es the te xture of the image in Figure 5(b4) and smooths the sk y in Figure 5(c4) more accurately as compared with WESNR. These results demonstrate the ef fecti v eness of the proposed procedure for updating the weights and to the replacement of AMF [24] adopted in WESNR by R OE [27] which in combination reduces the ef fect of hea vy tails of the impulse noise, especially at high impulse noise density rates. Figure 5. Denoising results of mix ed g aussian plus impulse noise Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1440 1450 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1449 7. CONCLUSION This paper proposes a denoising algorithm in order to reco v er images highly corrupted by mix ed Gaus- sian plus impulse noise. The promising results sho w that the proposed algorithm preserv es image fine details more accurately as compared with WESNR especially for highly corrupted images. The main contrib utions of the proposed algorithm are the emplo yment of a more accurate initialization image through replacing AMF by R OE and the use of a simple parameter -free procedure for updating encoding weights W by a v oiding the use of the e xponential function and the need for calibrating the decaying rate parameter ( c ) in accordance with the noise state le v el. As a quantitati v e e xample of the impro v ement in results, for =30 and sp =70 % , the proposed algorithm accomplishes an a v erage PSNR g ain of approximately 2.32dB o v er the entire images under test.The promising results of the proposed algorithm open the w ay for further impro v ement, enabling it to eliminate more comple x Gaussian plus random v alued impulse noise. This issue is left as a possible future w ork. REFERENCES [1] Y .-B. Zheng, T .-Z. Huang, X.-L. Zhao, T .-X. Jiang, T .-H. Ma, and T .-Y . Ji, “Mix ed noise remo v al in h yperspectral image via lo w-fibered-rank re gulariza tion, IEEE T r ansac tions on Geoscienc e and Remote Sensing , v ol. 58, no. 1, pp. 734-749, Jan. 2020, doi: 10.1109/TGRS.2019.2940534. [2] Y . Chen, J. Li, and Y . Zhou, “Hyperspectral image denoising by tota l v ariation-re gularized bilinea r f actorization, Signal Pr ocessing , v ol. 174, p. 107645, 2020, doi: 10.1016/j.sigpro.2020.107645. [3] J. Chen, G. Zhang, S. Xu, and H. Y u, A blind cnn denoising model for random-v alued i mpulse noise, IEEE Access , v ol. 7, pp. 124647-124661, 2019, doi: 10.1109/A CCESS.2019.2938799. [4] H. Zeng, X. Xie, W . K ong, S. Cui, and J. Ning, “Hyperspectral image denoising via combined non-local self- similarity and local lo w-rank re gularization, IEEE Access , v ol. 8, pp. 50190-50208, 2020, doi: 10.1109/A C- CESS.2020.2979809. [5] S. A. Hussain, A. A. Hussein, and A. H. Reja, “F ast and rob ust random-v alued ima ge denoising algorithm based on road statistics, Pr oceedings of 187th The IIER International Confer ence , Hong K ong , 2018, pp. 12-18. [6] A. A w ad, “Remo v al of fix ed-v alued impulse noise based on probability of e xistence of the image pix el, In- ternational J ournal of Electrical & Computer Engineering (IJECE) , v ol. 8, no. 4, 2018, pp. 2106 2114, doi: 10.11591/ijece.v8i4.pp2106-2114. [7] H. Zhu and M. K. Ng, “Structure d dictionary learning for image denoising under mix ed g aussian and impulse noise, IEEE T r ansactions on Ima g e Pr ocessing , v ol. 29, pp. 6680-6693, 2020, doi: 10.1109/TIP .2020.2992895. [8] V . S. T allaprag ada, N. A. Mang a, G. P . K umar , and M. V . Naresh, “Mix ed image denoising using weighted coding and non-local similarity , SN Applied Sciences , v ol. 2, no. 6, pp. 1–11, 2020, doi: 10.1007/s42452-020-2816-y . [9] J. Jiang, L. Zhang, and J. Y ang, “Mix ed noise remo v al by weighted encoding with sparse nonlocal re gularization, IEEE tr ansactions on ima g e pr ocessing , v ol. 23, no. 6, pp. 2651-2662, June 2014, doi: 10.1109/TIP .2014.2317985. [10] L. Li u, L. Chen, C. L. P . Chen, Y . Y . T ang and C. M. pun, “W eighted joint sparse representation for remo v- ing mix ed noise in image, IEEE tr ansactions on cybernetics , v ol. 47, no. 3, pp. 600-611, March 2017, doi: 10.1109/TCYB.2016.2521428. [11] L. Liu, C. P . Chen, X. Y ou, Y . Y . T ang, Y . Zhang, and S. Li, “Mix ed noise remo v al via rob ust constrained sparse representation, IEEE T r ansactions on Cir cuits and Systems for V ideo T ec hnolo gy , v ol. 28, no. 9, pp. 2177-2189, Sept. 2018, doi: 10.1109/TCSVT .2017.2722232. [12] J. Liu, X.-C. T ai, H. Huang, and Z. Huan, A weighted dictionary learning model for denoising images cor - rupted by mix ed noise, IEEE tr ansactions on ima g e pr ocessing , v ol. 22, no. 3, pp. 1108-1120, March 2013, doi: 10.1109/TIP .2012.2227766. [13] M. Filipo vi ´ c and A. Juki ´ c, “Restorat ion of images corrupted by mix ed g aussian-impulse noise by iterati v e soft-hard thresholding, in 2014 22nd Eur opean Signal Pr ocessing Confer ence (EUSIPCO) , 2014, pp. 1637-1641. [14] P . Zhuang, W . W ang, D. Zeng, and X. Ding, “Rob ust mix ed noise remo v al with non-parametric bayesian sparse outlier model, in 2014 IEEE 16th International W orkshop on Multimedia Signal Pr ocessing (MMSP) , 2014, pp. 1-5, doi: 10.1109/MMSP .2014.6958792. [15] P . Zhuang, Y . Huang, D. Zeng, and X. Ding, “Mix ed noise remo v al based on a no v el non-parametric bayesian sparse outlier model, Neur ocomputing , v ol. 174, pp. 858–865, 2016, doi: 10.1016/j.neucom.2015.09.095. [16] H. K. Agg arw al and A. Majumdar , “Mix ed g aussian and impulse denoising of h yperspectral images, in 2015 IEEE International Geoscience and Remote Sensi ng Symposium (IGARSS) , 2015, pp. 429-432, doi: 10.1109/IGARSS.2015.7325792. [17] H. K. Agg arw al and A. Majumdar , “Hyperspectral image denoising using spatio-spectral total v ariation, IEEE Geo- science and Remote Sensing Letter s , v ol. 13, no. 3, pp. 442-446, March 2016, doi: 10.1109/LGRS.2016.2518218. Ima g e mixed gaussian and impulse noise elimination based on spar se ... (Ahmed Abdulqader Hussein) Evaluation Warning : The document was created with Spire.PDF for Python.