Engineering, Technology & Applied Science Reearch Vol. 7, o. 6, 207, 2288-2292 2288 Random Valued Impule oie Removal Uing Region Baed Detection Approach Shubhendu Banerjee arula Intitute of Technology Kolkata Wet Bengal, India hankuhubhendu@gmail.com Aritra Bandyopadhyay Supreme Knowledge Foundation Group of Intitution Mankundu Wet Bengal, India aritra.d90@gmail.com Avik Mukherjee Tata Conultancy Service Ltd Digital Interactive IOU in TCS Pune, India mukherjee.avik852@gmail.com Atanu Da etaji Subhah Engineering College Garia, Kolkata Wet Bengal, India atanuda75@yahoo.co.in Rajib Bag Supreme Knowledge Foundation Group of Intitution Mankundu Wet Bengal, India rajib.bag@gmail.com Abtract Removal of random valued noiy pixel i extremely challenging when the noie denity i above 50%. The exiting filter are generally not capable of eliminating uch noie when denity i above 70%. In thi paper a region wie denity baed detection algorithm for random valued impule noie ha been propoed. On the bai of the intenity value, the pixel of a particular window are orted and then tored into four region. The higher denity baed region i conidered for tepwie detection of noiy pixel. A a reult of thi detection cheme a maximum of 75% of noiy pixel can be detected. For thi purpoe thi paper propoe a unique noie removal algorithm. It wa experimentally proved that the propoed algorithm not only perform exceptionally when it come to viual qualitative judgment of tandard image but alo thi filter combination outmart the exiting algorithm in term of MSE, PSR and SSIM comparion even up to 70% noie denity level. Keyword-random valued inpule noie; image filtering; region baed; detection I. ITRODUCTIO Random valued impule noie removal i a tringent tak. Thi type of noie i introduced in the digital image through tranmiion and acquiition []. A ignificant characteritic of thi type of noie i that only fraction of the pixel are degraded [2]. A variety of filter [4-22] are being propoed in pat year to remove random valued impule noie. The main motive of uing thee filter i to detect and reduce the noie a well a preerving the image information. Median filter were rapidly ued to remove thi type of noie. Author in [] introduced a median filter which wa able to contain noie up to ome level. It wa a good filter, which preerved edge greatly. But thi filter wa not fulty effective a noiele pixel alo got changed during retoration. Succeively ome other median filter were propoed [4-6]. Thee filter provided noie retoration to a medium level but they were not o effective for the ame reaon, o the image got blurred and edge were modified. Afterward a variety of condition baed filter were introduced. But thoe filter failed to retore the noie effectively. Author in [7] introduced a two phae wap filter which wa able to preerve detail at a low noie denity. A tri tate median filter that ued a tandard median and center weighted median filter to detect the noie but wa not able to preerve minute detail at medium noie denitie wa propoed in [8]. In [9], author propoed another adaptive center weighted median filter which ighted a light improvement in preerving detail but that alo in lower noie denitie. Author in [0] introduced a filter which wa able to preerve minute detail at medium noie denity. Though the detection cabability wa not that impreive. To enhance it capability a generic programming filter wa propoed in []. Thi filter had a two tage cacading detector which enchanced the detection rate. A new dimenion in the filtering i introduced in [2]. Thi wa a PDE baed technique which ued aniotropic diffuion to filter impulive noie. The method performed better than the previou dicued cheme. In 202 ROR [] (robut outlyingne ratio) wa propoed. Thi filter firt calucated a robut outlyingne ratio to find out the impulive pixel then by the ROR value the pixel are divide into four cluter. It wa followed by a two tage detection proce. Ue of fuzzy filter had alo hown great reult. In 20 a robut direction baed detector [4] wa propoed which ued tandard deviation baed detection procedure to filter the noie. Afterward, TDWM [5] wa propoed. Thi filter ued directional threhold baed approach to remove noie. Thi filter wa good at medium noie denity level. Recently, a new detection method wa introduced in [6]. Thi method ued tandard deviation and threhold concept together to approximate neighborhood pixel. Thi filter wa ucceful at 60% noie denity but failed once the noie wa higher than that. So, removal of impule at higher noie denity wa till a
Engineering, Technology & Applied Science Reearch Vol. 7, o. 6, 207, 2288-2292 2289 tak to be achieved. In thi paper, a two-tep, region baed, detection and removal method i introduced. Thi method generated improved reult in higher noie denitie. The ret of the paper i organized a follow. Propoed methodology i dicued in Section II. Computer-generated reult and comparion with recent filter are preented in Section III. Finally future challenge and concluion are dicued in Section IV. II. PROPOSED METHODOLOGY Let x i,j for (i,j) Є A {,2,, M} {,2,, } be the intenity of the pixel at pixel location (i,j) of a random valued corrupted M image X. Simultaneouly a ame ize binary flag image F ha been generated taking all value of the pixel f i,j =. A. oie Detection Part A 5 5 window (W) ha been created taking center pixel in x i,j the image X. A et {S} i formed taking all the element of the window and the element are put in increaing order. S= {, 2,, 4,., 25} 2 25 Four ubet S, S2, S, S4 are alo formed like below. S= { i : 0 i 60 } i =,2,,4,..,25 S2= { i : 6 i 20 } i =,2,,4,..,25 S= { i : 2 i 80 } i =,2,,4,..,25 S4= { i : 8 i 255 } i =,2,,4,..,25 ow n(s), n(s2), n(s), n(s4) are calculated. Examining Max n(si) i=,2,,4 the following tep are followed: if n(s) i Max m 7 6 8 ele if n(s2) i Max m 0 9 ele if n(s) i Max ele m m 2 4 7 8 9 Hence tandard deviation (σ) i calculated taking all the element of W with equation no. 2 ( ) () i x i if i j x j m i, then x, i conidered a a unditurbed pixel ele noiy pixel and imultaneouly pixel of the binary image F i replaced by 0 i.e f 0. i, j After completing the above proce the updated binary flag image F i generated. The above algorithm i baed on maximum no of element of the four ubet. Here it i conidered that the unditurbed pixel will increae the no of element in the correponding ubet. Here S i a et of element in increaing order. The unditurbed pixel will take their poition 6,7,8 for S, 9,0, for S2, 2,,4 for S and 7,8,9 for S4 repectively. A the aforementioned algorithm i not ufficient to find out the diturbed element above 75%, for final detection and filtering, next algorithm are followed. B. oie Removal Part The image X (M ) and the binary flag image F (M ) are taken for further operation. Here x i,j and f i,j are conidered a pixel value at the location i, j in the image X and F. if (f i,j == 0) then replace x i,j = 0. Completing the above proce image X (M ) i generated. ) Algorithm if (x i,j == 0) then ue SCMMF [7] algorithm for removal proce. Completing all the tep ame ize image X6 (M ) i generated 2) Algorithm 2 A matrix from image X6 i created taking x i,j a center. Here x i,j and f i,j are conidered a pixel value at the location i,j in the image X6 and F. if (f i,j == ) then Let A = {x : x 0} replace Let B = { y : y = if (n(b) 5) then y x i,j= n(b) x i, j x σ} Completing all the tep a final ame ize output image X7 (M ) i generated. III. RESULT AD AALYSIS Thi method i evaluated uing MATLAB 7.2 (R20a) in a 2.8-GHz CPU with 2 GB RAM. To check the effetene of
Engineering, Technology & Applied Science Reearch the propoed algorithm, random Matlab image are ued. oie denity i varied from 0% to 70%. The quality analyi of the propoed algorithm i hown in Figure. Vol. 7, o. 6, 207, 2288-2292 2290 (PSR), Structural Similarity Index Meaurement (SSIM) [8] and Mean Structural Similarity Index Meaurement (MSSIM) a clarified in (2), (), (4), and (5) repectively: (m, n) O (m, n) O M M, MSE= 2 (2) where, O = Original Image ^ O = De-noied image M= umber of row = umber of column PSR 0log0 SSIM( x, y) 2552 MSE () (2 x y C )(2 xy C2 ) (4) ( x2 y2 C )( x2 y2 C2 ) M (5) SSIM(xm, ym ) M m where μx and μy are the mean of image x and image y repectively. The tandard deviation of image x and image y i denoted by σx and σy repectively. C, C2 are the contant σxy i the co-variance of x and y, given by : MSSIM( x, y) (d) (e) x x i i 2 x ( xi x ) 2 i i xy ( xi x ) i i (f) (g) (h) (i) Fig.. Reult of propoed filter on Lena image Original Image 0% noie corrupted image propoed filter output (0%) (d) 40% noie corrupted image (e) propoed filter output (40%) (f) 50% noie corrupted image (g) propoed filter output (50%) (h) 60% noie corrupted image (i) propoed filter output (60%) The comparion of PSR value of the propoed algorithm with repect to known filter i repreented in Table I. At high noie denity level i.e. 40-60% the propoed method obviouly outperform the exiting filter. Table II repreent the PSR value of different image at 70% noie denity. Figure 2 tate the perceptible reult at image Lena, Gold Hill and Cameraman at 70% noie denity level. MSSIM i ued to evaluate the tructural ymmetry of Lena image at varying noie denitie, hown in Figure 4. The PSR value of the popular image like Lena, Cameraman and Gold Hill at 70% noie denity with repect to the propoed method are dicued in Table II. From the qualitative, numeric and tructural analyi it i obviou that the performance of the propoed method produced uperior performance than other recent filter epecially at high noie denity level. Figure reflect the comparion of the propoed filter with other variou filter by the graph of MSE againt the noie denity 0%-60% for the Lena image. umeric analyi of the propoed method i quantified by Mean Square Error (MSE), Peak Signal-to-oie Ratio
Engineering, Technology & Applied Science Reearch Vol. 7, o. 6, 207, 2288-2292 229 TABLE I. COMPARISO OF PSR VALUES OF DIFFERET FILTERS FOR LEA IMAGE AT VARIOUS OISE DESITIES Image Filter 0% 40% 50% 60% LUO 0.29 28.27 26.2 24.2 GP.87 28.42 24.86 2.68 SDD 0.86 28.5 26.9 24.08 Lena ROR 24.9 2.5 8.69 5.6 RDD 26.4 22.98 9.99 6.72 TDWM 28.02 26.85 25.88 25.00 [6] 0.68 29.4 27.4 25.0 Propoed 29.80 29.45 28.64 28.49 The tudied filter LUO, GP, SDD, ROR, RDD, TDWM are chiefly explored and a comparative tudy with pecial emphai at high noie denity level i preented. Earlier contribution in thi area of work are revied, re-embellihed and produced in a more comprehenive but implified manner that would make thi work undertandable even to people with general cience background. Thi work focue on the exiting renowned filter and alo on the latet finding. It wa oberved that the uage of thoe filter wa not up to the mark at high denity noie level. Thi work i an attempt to bridge that gap by introducing the modified filter. TABLE II. PSR FOR DIFFERET IMAGES FOR PROPOSED METHOD AT 70% OISE DESITY PSR Lena Gold Hill Cameraman 26.0 26.7 25.47 (d) Fig. 2. Reult of different image by propoed filter at 70% noie denity Lena Gold-hill Cameraman. Fig. 4. SSIM Index map (MSSIM) of Lena image at noie denity: 0% (MSSIM=0.8775), 40% (MSSIM=0.869), 50% (MSSIM=0.8567), (d) 60% (MSSIM= 0.8542). Fig.. Image. Comparion graph of MSE at different noie denitie for Lena IV. COCLUSIOS Thi paper take into account both traditional and avantgarde random valued impulive noie removal filter technique. REFERECES [] S. Bandyopadhyay, S. Banerjee, A. Da, R. Bag, A Relook and Renovation over State-of- Art Salt and Pepper oie Removal Technique, I.J. Image, Graphic and Signal Proceing, Vol. 9, pp. 6-69, 205 [2] Y. Dong, R. H. Chan, S. Xu, A detection tatitic for random valued impule noie, IEEE Tranaction on Image Proceing Vol. 6, o. 4, pp. 2-20, 2007 [] J. W. Tukey, Exploratory Data Analyi, Reading, Addiion- Weley, 97 [4] G. R. Arce, R. E. Foter, Detail-preerving ranked-order baed filter for image proceing, IEEE Tranaction on Acoutic, Speech, and Signal Proceing, Vol. 7, o., pp. 8 98, 987 [5] W. Y. Han, J. C. Lin, Minimum-maximum excluive mean (MMEM) filter to remove impule noie from highly corrupted image, Electronic Letter, Vol., o. 2, pp. 24 25, 997 [6] Y. H. Lee, S. A. Kaam, Generalized median filtering and related nonlinear filtering technique, IEEE Tranaction on Acoutic, Speech, and Signal Proceing, Vol., o., pp. 672 68, 985 [7] E. Abreu, M. Lighttone, S. K. Mitra, K. Arakawa, A ew Efficient Approach for the Removal of Impule oie from
Engineering, Technology & Applied Science Reearch Vol. 7, o. 6, 207, 2288-2292 2292 Highly Corrupted Image, IEEE Tranaction on Image Proceing, Vol. 5, o. 6, pp.02-025, 996 [8] T. Chen, K. K. Ma, H. L. Chen, Tri-tate median-baed filter in image de-noiing, IEEE Tranaction on Image Proceing Vol. 8, o. 2, pp. 84 88, 999 [9] T. Chen, H R Wu, Adaptive impule detection uing center weighted median filter, IEEE Signal Proceing Letter, Vol. 8, o., pp., 200 [0] W. Luo, An efficient algorithm for the removal of impule noie from corrupted image AEU - International Journal of Electronic and Communication, Vol. 6, o. 8, pp. 55 555, 2007 []. I. Petrovic, V. Crnojevic, Univeral impule noie filter baed on genetic programming, IEEE Tranaction on Image Proceing, Vol. 7, o. 7, pp. 09 20, 2008 [2] J. Wu, C. Tang, PDE-Baed Random-Valued Impule oie Removal Baed on ew Cla of Controlling Function, IEEE Tranaction on Image Proceing, Vol. 20, o. 9, pp. 2428-248, 20 [] B. Xiong, Z. Yin, A Univeral de-noiing framework with a new impule detector and nonlinear mean, IEEE Tranaction on Image Proceing, Vol. 2, o. 4, pp. 66-675, 202 [4] K. Prathiba, R. Rathi, C. S. Chritopher, Random Valued Impule Denoiing uing Robut Direction baed Detection, IEEE Conference on Information and Communication Technologie, pp. 27-242, 20 [5] Sarkar, S. Changder, J. K. Mandal, A Threhold baed Directional Weighted Median Filter for Removal of Random Impule in Thermal Image, IEEE 2nd International Conference on Buine and Information Management, pp. 69-74, 204 [6] S. Banerjee, A. Bandyopadhyay, R. Bag, A. Da, A Deviation Baed Identification of Random Valued Impule oie Toward Image Filtering Uing eighborhood Approximation, rd International Conference, Foundation and Frontier in Computer, Communication and Electrical Engineering, pp. 27-220, 206 [7] S. Banerjee, A. Bandyopadhyay, R. Bag, A. Da, Sequentially Combined Mean-Median Filter for High Denity Salt and Pepper oie Removal, IEEE International Conference on Reearch in Computational Intelligence and Communication etwork, pp. 2-26, 205 [8] Z. Wang, C. A. Bovik, R. H. Sheikh, P. E. Simoncelli, Image quality aement from error viibility to tructural imilarity, IEEE Tranaction on Image Proceing, Vol. 4 o., pp. 600-62, 2004 [9] Johi, A. K. Boyat, B. K. Johi, Impact of Wavelet Tranform and Median Filtering on Removal of Salt and Pepper oie in Digital Image, International Conference on Iue and Challenge in Intelligent Computing Technique, pp. 88-84, 204 [20] S. Banerjee, A. Bandyopadhyay, R. Bag, A. Da, Moderate Denity Salt & Pepper oie Removal, International Journal of Electronic & Communication Technology, Vol. 6, o., pp. 44-48, 205 [2] S. Banerjee, A. Bandyopadhyay, R. Bag, A. Da, eighborhood Baed Pixel Approximation for High Level Salt and Pepper oie Removal, CiiT International Journal of Digital Image Proceing, Vol. 6, o. 8, pp. 46-5, 204 [22] S. Banerjee, A. Taraphdar, R. Bag, A. Da, Binary expanion baed de-noiing algorithm for an image corrupted by Gauian noie, CSI Tranaction on ICT, Springer, pp. -5, 206 AUTHORS PROFILE Shubhendu Banerjee wa born in 988, received hi B.Tech (Computer Sc) &, M.Tech (Computer Sc) both from Wet Bengal Univerity of Technology, India in the year of 200 & 202 repectively. Preently, he i working a an Aitant Profeor in the department of Computer Science & Engineering at arula Intitute of Technology under Wet Bengal Univerity Technology, Agarpara, Kolkata, Wet Bengal, India. He ha 9 publication in International Journal, and Conference proceeding to hi credit. Hi reearch interet include image and ignal proceing. Aritra Bandyopadhyay wa born in 988, received hi B.Tech (Computer Sc) & M.Tech (Computer Sc) both from Wet Bengal Univerity of Technology, India in the year of 200 & 202 repectively. Preently, he i working a an Aitant Profeor in the department of Computer Science & Engineering at Supreme Knowledge Foundation Group of Intitution under Wet Bengal Univerity Technology, Mankundu, Hooghly, Wet Bengal, India ince 202. He ha 9 publication in International Journal to hi credit. Hi reearch interet include image and ignal proceing. Avik Mukherjee wa born in 99, received hi B.Tech (Computer Sc) from Wet Bengal Univerity of Technology, India in the year of 205. Preently, he i working a Sytem Engineer (Developer) in the Digital Interactive IOU in TCS, Pune. He ha hand on experience in working in java, Android development, Java cript, jquery & plugin, Boottrap, Angularj and other front end framework. Dr. Atanu Da wa born in 975, received hi BSc (Math Hon.), MSc (Statitic) degree from The Univerity of Burdwan and ME & PhD (Engg) degree from Jadavpur Univerity, India in the year 996, 998 and 2002 & 20 repectively. Hi doctoral work wa in the field of etimation and filtering. He i working a an Aitant Profeor, CSE at etaji Subhah Engineering College under Wet Bengal Univerity Technology, Kolkata, India ince 2002. He ha more than 20 publication in reputed refereed journal, edited volume and conference proceeding. Hi reearch interet include image and multimedia proceing & education technology beide etimation and filtering technique for evolving ytem. Dr. Rajib Bag wa born in 969, received hi B.Sc (Phyic Hon.) from Calcutta Univerity, M.Sc. (Phyic) from Vinoba Bhave Univerity and M.Tech. & Ph.D (Engg.) from Jadavpur Univerity, India in the year of 99, 996, 2007 & 202 repectively. Hi doctoral work wa in the field of control ytem. Preently, he i working a a Profeor & Head in the department of Computer Science & Engineering at Supreme Knowledge Foundation Group of Intitution under Wet Bengal Univerity Technology, Mankundu, Hooghly, Wet Bengal, India. He ha more than 20 publication in reputed refereed journal and conference proceeding to hi credit. Hi reearch interet include image and ignal proceing & education technology beide control ytem