STATISTICAL COMPLEXION-BASED FILTERING FOR REMOVAL OF IMPULSE NOISE IN COLOR IMAGES

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 STATISTICAL COMPLEXION-BASED FILTERING FOR REMOVAL OF IMPULSE NOISE IN COLOR IMAGES 1 P.VENKATESAN, 2 Dr.S.K.SRIVATSA 1 Senior Assistnt Professor, Deprtment of Electronics nd Communiction Engineering, SCSVMV University, Knchipurm, Tmil Ndu, Indi. 2 Senior Professor, St. Joseph s College of Engg, Jeppir Ngr, Chenni-600 E-mil: 1 pv.eceknchi@gmil.com, 2 profsks@rediffmil.com ABSTRACT A sttisticl complexion-bsed filtering techniques, nmed s the Adptive Sttisticl Complexion bsed Filtering techniques (ASCF), is presented for removl of impulse noise in degrded color imges. In distinction with the trditionl noise detection techniques where only 1-D numericl informtion is used for noise detection nd estimtion, n innovtive noise detection scheme is proposed bsed on sttisticl personlity nd fetures (i.e., the 2-D informtion) of the degrded pixel or the pixel region, leding to effective nd efficient noise detection nd estimtion outcomes. A progressive restortion mechnism is devised using multipss nonliner opertions which dpt to the intensity nd the types of the noise. widespred experiments conducted using extensive rnge of test color imges hve shown tht the ASCF is dvnced to number of existing well-known stndrd techniques, in terms of verge imge restortion performnce criteri, including objective mesurements, the visul imge qulity, nd the computtionl complexity. Keywords: color imge restortion, impulse noise detection, progressive filtering. 1. INTRODUCTION Imges re often corrupted by impulse noise due to fculty imge cquisition device or to chnnel trnsmission errors, much reserch hs been done on removing such noise. The noise objective is to suppress the noise while preserving the integrity of edges nd detil informtion. To this end, nonliner methods hve been found to provide more stisfctory results thn liner techniques. The most frequently used nonliner method it the medin filter [Arc86], which is superior to liner filters in its bility to suppress impulse noise nd preserve edges. Nonliner filtering techniques hve been extensively reserched in the lst decde due to their effectiveness in restortion of impulse noise degrded color imges. The medin filter is usully used to remove impulse noise. Compred with liner filters, the medin filter is more powerful in tht single corrupt or noisy pixel in the filtering window will not ffect the medin vlue extensively. For removl of noise in color imges, vrious vector medin filters hve proven relevnt nd effective. Amongst the erly publictions, the most well known vector filters for color imge denoising include the vector medin filter (VMF), the vector directionl filter (VDF), nd the directionl distnce filter (DDF). While these vector filters perform well in suppressing the 533 impulse noise, they introduce imge distortions such s blurring round edges nd in detil res which feture high sptil frequency contents nd vritions. Different types of weighted nonliner filtering techniques hve been investigted over the yers to chieve better performnce in both noise suppression nd detil preservtion. Recently, fuzzy filtering techniques hve been developed, chieving powerful imge denoising performnce. A clss of chromtic filters for imge restortion in the color spce ws lso proposed to chieve better chromtic smoothness. Adptive filters hve demonstrted their effectiveness in imge restortion considering vrious types of noise with different distributions nd imge structures. In shrp contrst with the dditive noise tht contmintes ll imge pixels, the impulse noise destroys only some portion of n imge nd leves other pixels noise-free. Detection bsed vector filtering techniques such s the dptive vector medin filter (AVMF), the dptive vector LUM(lower-upper middle) smoother (AVLUM), modified weighted vector medin filter (MWVM), nd the dptive selection center weighted vector direction filter (ACWVDF) were specilly designed to remove the impulse noise from color imges. They utilize series of weighted medin vector filters to perform binry noise

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 detection nd switch between the output of n identity filter nd tht of weighted medin vector filter, ccording to the detection results. A survey of nonliner vector filtering ws presented in for impulse noise removl from color imges. 2. STATISTICAL MODELS OF IMPULSE NOISE Color imges my be contminted by vrious types of noise nd impulse noise is the noise model frequently used nd reported in digitl restortion literture. Impulse noise corruption often occurs in digitl imge cquisition or trnsmission process s result of photo-electronic sensor fults or chnnel bit errors. Imge trnsmission noise my be cused by vrious sources, such s cr ignition systems, industril mchines in the vicinity of the receiver, switching trnsients in power lines, lightning in the tmosphere nd vrious unprotected switches. This type of trnsmission noise is often modeled s the impulse noise. The impulse noise cn lso be introduced into imges during cquisition of the imges. For exmple, the impulse noise my be introduced during fingerprint cquisition in rel-life border security check. For more bckground informtion bout the physicl model of the impulse noise, we refer reders to. The two most common impulse noise types re fixed-vlue impulse noise (lso known s the slt-nd-pepper noise) nd rndom-vlue impulse noise. Let C = {c = (c 1,c 2 ) 1 c 1 H,1 c 2 W} denote the set of the pixel coordintes of color imge, where H nd W re height nd the width of the imge, respectively t ech pixel coordinte c C, multivrite vlue vector in the RGB color spce, X(C) = [x R (c), x G (c), x B (c)] T, is used to represent the RGB(Red,Green, Blue) pixels vlues. Two pproches s reported in the literture re used in this pper to model the impulse noise for color imge restortion. In the first pproch, the impulse noise corruption of the color imges in the RGB spce is expressed by multivrite model. Y(c) = s(c), with probbility (1-P I ) n T (c), with probbility P I (1) nd in the second pproch the impulse noise corruption of the color imges in the RGB spce is expressed by multivrite model. Y(c) = s(c), with probbility (1-P) 3 n t (c), with probbility 1-(1-P) 3 (2) Where S(c) nd X(c) represent the originl nd the observed pixel (vector) vlues t coordinte c, respectively, nd the vlue of n T (c) nd n t (c) is generted by substituting t lest one color component of the pixel S(c) by distinct vlue d in both (1) nd (2). In (1), P I is the impulse noise rtio; fctor r=0.5 is used to simulte the chnnel correltion for ech corrupted pixel, nmely if t lest one of the three components of the pixel is corrupted by the impulse noise, its remining noise free components will hve 50% probbility to be corrupted. The second pproch (2) is more generlized impulse noise model of color imges where P = P R = P G = P B is the impulse noise rtio for ech chnnel of corrupted color imge, ssuming tht the imge is corrupted by the impulse noise in chnnel independent mnner. In (1) nd (2),if d, the component vlue of n t (c) or n T (c) equls the mximum or the minimum vlue of the digitl imge (e.gg,, 255 or 0 for n 8- bit chnnel of the 24- bit color imge in the RGB spce), the impulse noise is referred to s the slt nd pepper impulse. Ech pixel of the imge my be corrupted by either the pepper or slt impulse with unequl probbilities. However, if the mplitudes of the impulse re distributed rndomly with, e.gg, the uniform or the Gussin distribution, in the rnge of [0,255], more generl type of the impulse noise is generted nd nmed s the rndom impulse noise. The impulse noise cn be represented by joint probbility distribution describing the sptil distribution of the impulses s well s their mplitudes. As is typiclly the cse, these two quntities re considered to be independent. In this pper, n Adptive Sttisticl Complexion bsed Filtering techniques (ASCF) with low computtionl complexity is proposed for restortion of digitl color imges corrupted by the impulse noise. This technique uses set of novel noise detection criteri for detection of the corrupted pixels, which re bsed on 2-D geometric nd dimension fetures of the noisy pixel or the noisy region of imges. This is in contrst with the trditionl noise detection techniques where only 1-D sttisticl informtion is used for estimtion of the noise rtio nd the noise sttisticl distribution model. Bsed on the result of the estimtion, n dptive progressive filtering opertion is employed in combintion with 534

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 optimized dimension nd shpe of processing windows.computtionl efficiency of the ASCF is lso investigted.denoising performnce of the ASCF is evluted to demonstrte noticeble gins ginst tht of number of well-known benchmrk techniques mentioned bove, in terms of stndrd objective mesurements perceptul imge qulity nd computtionl complexity, especilly for suppression of the impulse noise in medium-nd lrge-size color imges 3. DIMENSIONAL AND GEOMETRIC FEATURES OF IMPULSE NOISE A mjor problem in restortion of color imges to dte is the destruction of detiled imge structures due to inbility of denoising filters to distinguish cluster of corrupted pixels from cluster of pixels presenting fine (detiled) imge structures nd the incorrect removl or modifiction of pixel segments. This section describes novel technique which detects nd removes, effectively nd efficiently, impulses in color imges. As defined in, ny two pixels t (i 1, j1) nd (i 2, j2) re clled 4-neighbors, if they hve city block distnce D4=1 from ech other. Similrly, 8- neighbors re two pixels with chessbord distnce D8=1. The city block distnce is defined s D4 ((i 1, j1), (i 2,j 2 )) = i 1 - i 2 + j 1 - j 2 nd the chessbord distnce is defined s D8 ( (i 1, j 1 ), (i 2,j 2 ) ) = Mx { i 1 -i 2, j 1 - j 2 }. For exmple, ech color imge pixel in fig.1(d) is represented by the coordintes of its position, i.e., pir of integers (column number, row number ).given pixel ( 3,3 ), for instnce, its 4-neighbors re (2,3),(3,4),(3,2), nd (4,3) nd its 8- neighbors re its 4-neighbors plus (2,2),(4,4),(2,4) nd (4,2). Creful exmintion of vriety of color imges corrupted by the fore mentioned impulse noise models revels tht most of uncorrupted pixels or pixel regions in nturl color imge demonstrte certin degree of smoothness. This mens tht the color intensities of pixel lwys chnge grdully in ll its 8-neighbors directions (e.g., in smooth re), or chnge grdully t lest in one (edge) direction (e.g., in boundry re). In contrst with norml or uncorrupted pixels of imges, impulse noise corrupted pixels lwys stnd out s n isolte spot or cluster by its very un-hrmonious colors, shpes nd sizes compred with those of its neighborhood. Even in the boundry (or edge) re, uncorrupted objects in nturl color imges hve different types of edges from those corrupted by the impulses. It is observed tht lmost ll impulses only hve shrp step edges nd, in contrst lmost none of the uncorrupted objects hve this type of edges in its vicinity. The borders of the uncorrupted objects still hve nrrow trnsitionl region of few pixels, even in the grdient direction of shrp chnging boundry re. In cses where imges corrupted by the impulse with the low noise rtio, the sizes of the corrupted pixels (i.e., corrupted res) re most likely represented by isolted individul pixels or short line with one pixel width. The pixels of the line my be djcent in the digonl direction.with the increse of the noise rtio, corrupted pixel regions /clumps with two pixel width in two perpendiculr directions my occur long with the individul impulses nd smller impulse regions s shown in imges corrupted with the low impulse rtio. The shpes of the noise regions my be isolted point, short thin line, cross of two short thin lines or other smll round shped blocks. In other words, with the increse of the noise rtio, the noise my pper isolted or clustered with more different sizes nd shpes. According to the bove observtions nd nlysis of color, shpes nd sizes of impulse noise corrupted pixels /regions, nd the types of edges which form the borders of the noise regions, novel impulse noise detection method is devised here bsed on 2 D geometric fetures of the impulses, insted of the 1-D rnk ordered sttisticl informtion used by other well know filtering techniques, to determine if ech pixel in color imge is corrupted or clened. One of the geometric properties of the impulse noise is the edge feture of its boundry. An edge cn be defined s locl discontinuity in color component or illumintion intensity function nd the edge orienttion is defined s edges of n octgonlly shped object whose mplitude is higher or lower thn its bckground. Therefore, the criteri for identifying the edge feture round the pixel re bsed on the two types of derivtives, which re pproximted by pixel differences in digitl color imges. Given tht Y(c) = [ y R (c), y G (c), y B (c) ] T is the vector contining color components functions of color imge, the two specil types of prtil derivtives re denoted s Y(c)/ c nd Y(c)/ c d respectively. Y(c)/ c t C= (i, j) is pproximted by G, the difference between the pixel nd its 4- neighbors for ech component of the color pixel, nd defined s follows: 535

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 S 1 (n 1 ) = y [i.j]-y [i-n 1,j] S 2 (n 2 ) = y [i,j]-y[i, j-n 2 ] S 3 (n 3 ) = y [i,j]-y[i+n 3,j] S 4 (n 4 ) = y [i,j]-y[i, j+n 4 ] [3] Where n = [n 1,n 2,n 3,n 4 ] T, n k > 0 nd the defult vlue of n k is 1,for 1 k 4, nd subscript T represents the trnspose opertion. When derivtive is only considered in the digonl direction, y(c)/ c d is pproximted by G d, the difference between the pixel nd its other 8- neighbors, for ech component of the color component, nd defined s follows: G d 1 (n d 1 ) = y[ i,j]-y[ i-n d 1, j-n d 1 ] G d 2 (n d 2 ) = y[ i,j]-y[ i+n d 2, j-n d 2 ] G d 3 (n d 3 ) = y[ i,j]-y[ i+n d 3, j+n d 3 ] G d 4 (n d 4 ) = y[ i,j]-y[ i-n d 4, j+n d 4 ] [4] Where n d = [n d 1, n d 2, n d 3, n d 4 ] T d, n k > 0 nd the k defult vlue of n d is 1, for 1 k 4. The two specil derivtives, G nd G d, will be used to mesure the edge feture (shrpness) nd other geometric properties to determine whether center pixel t c = (i,j) is corrupted or not in the ASCF technique. 4. PRINCIPLE OF ASCF TECHNIQUE In detecting nd removing impulse noise filter my mke three min types of mistkes. Type I error (miss) occurs when there is corrupted pixel which the filter does not detect. Type II error (flse lrm) hppens when the filter detects n impulse noise pixel which is ctully clen. When the filter removes n impulse noise nd replces it with vlue determined by certin restortion strtegy, type III error (over or under-correcting error) is defined s the difference between the resultnt vlue fter the restortion process nd true pixel vlue s the noise -free pixel ws.different types of the so-clled switching filters nd fuzzy filtering techniques hve been developed over the yers, chieving good performnce in both noise suppression nd detil preservtion. Similr to the other well-known benchmrk techniques including the so-clled switching filters. And fuzzy-bsed filtering techniques. The ASCF technique described in this section consists of two components, i.e., impulse detection nd impulse removl. The novel criteri used by the ASCF for noisy pixel detection re bsed on combintion of the 2-D edge, geometric nd size fetures of the noisy pixel/region in the imges. They deprt from trditionl noise detection techniques used by the other existing filters, which only use some properties of the edge of noisy pixel re 1-D rnk ordered sttisticl informtion round the noisy pixel. For exmple, multiple threshold frmework nd corruption detectors re used in bsed on sttisticl informtion bout the neighborhood of ech locl pixel to locte impulse noise nd to preserve clen pixels. Time-consuming multiple reference filtering nd complex prmeter trining process highly limit the usge of these filters in rel-time pplictions. The new criteri presented in this pper lso deprt from recently developed fuzzy impulse noise filtering techniques. For exmple, the fuzzy noise detection method is minly bsed on clcultion of fuzzy grdient vlues nd fuzzy resoning, nd the fuzzy membership function representing the impulse noise is simplifiction of the obtined noise histogrm. 5. A TWO DIMENSIONAL IMPULSE NOISE DETECTION A key component of the AGFF technique is novel impulse detection scheme bsed on the 2-D geometric informtion of the corrupted pixels. First, we define the edge feture identifiction threshold, T e, which represents the vlue of derivtive to distinguish the shrp step edges from other types of edges [3]. Since very short thin lines usully form impulse noise pixels, the length of line is lso used s feture to distinguish short noise line from fine line in color imges. The length threshold, T l, my be defined ccordingly to the noise rtio. Second, in terms of the pixel coordintes of color imge, C, set of corrupted pixels is defined s S 1 ={ c (( S < (-T e ))^ (S <(=T e ))) V((S >T e )^ (S >T e )), n d k N d, n k =1 [5] For 1 k 4, N d = {1, 2, 3 T m} T m = (T l +1)/2} Where T m is used to define corrupted pixel-sizes in nd its defult vlue is 2. According to (5), IF the two prtil derivtives S nd S d of pixel hve the sme sign while their mgnitude re greter d thn preset threshold, when n k is 1 nd n k is 1 or 2 or 3, or (its components my hve different vlues), THEN the pixel belongs to. Set includes individul impulse pixels, slnt noise lines with one-pixel width nd the pixels of the lines only 536

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 djcent to ech other in digonl direction within the defined length of Tl. Third, set of corrupted pixels, which include individul impulse pixels, stright noise lines with one-pixel width the pixels of the lines being only 4-neighbors to ech other within the defined length of, is defined s S 2 ={ c (( S < (-T e ))^ (S d <(=T e ))) V ((G >T e ) ^ (G d >T e) ), n k N,, n k d =1 For 1 k 4, N = {1, 2, 3 T m} T m = (T l +1)/2} [6] Where Tm is used to define corrupted pixel-sizes in S 1 nd its defult vlue is 2. According to (6), IF the two prtil derivtives S nd S d of pixel hve the sme sign while their mgnitudes re greter thn preset Threshold T e,, when n d k is 1 nd n k is 1 or 2 or 3 Tm(its k components my hve different vlues), then the pixel belongs to S2. Next, set of corrupted pixels is defined s the S3, which include noisy pixels/regions within 3-pixel width in ny region except noisy pixels lredy in S1 nd/or S2, i.e., c S1US2. If S ={ c (( S < (-T e ))^ (S d <(=Te ))) V((S >Te )^ (S >Te )), n d k N d, n k =L} [7] Where 1 k 4, L is 2 or 3, nd the defult vlue for n d nd n is 2. Thus, S 3 cn be represented s S 3 = S - (S 1 US 2 ) [8] Where T m =2 for S 1 nd S 2 in (8). According to (8), IF the two prtil derivtives S nd S d of pixel hve the sme sign while their mgnitudes re greter thn preset threshold T e, when n d k nd n k re 2 or 3 nd the pixel is not in S 1 or S 2, THEN the pixel belongs to S 3. Finlly, ccording to observtion nd nlysis of vriety of nturl imges corrupted with the impulse noise, protrusive point in border re with high possibility of being corrupted pixel is defined s: S 4 = {c (( S k < (-T e ))^ (S d v <(=Te ))) V ((\S k >Te) ^ (S k >Te)), n d k = n k =n d v =, n v =1, V K {y y N n ^ y e}, e N n, N n = {1, 2, 3, 4}, (v=k) ^ (v {2, 3} V v {3, 4} V v {4, 1} V v {1, 2})}. [9] According to (9), IF the two prtil derivtives G d nd G d pixel hve the sme sign while their mgnitudes re greter thn preset threshold T e, with the prtil derivtives indexed by k contining only three out of the four distnce settings, nd the prtil derivtives indexed by being either {2, 3} or {3, 4} or{4, 1} or {1, 2} nd equl to k, when n d k,n k,n d v nd n v,,, nd re 1, THEN the pixel belongs to S 4. Since n impulse noise rtio p 1 <, U 1=1 n S i c C, where n=4, in the current design, T e in (5) nd (9) my be set t different vlues. The strtegy of the progressive restortion for the ASCF is, first, to restore corrupted individul pixels or noise regions of smll size. If it mde either Type II or Type III errors, it should not introduce ny new impulse noise regions bigger thn the existing ones. Then, further opertions re crried out round lrge noise corrupted regions to restore res of the imges ssocited with noise regions of the considerble size rebility In order to tke the dvntge of the medin filter nd to void the drw bcks. (I.e cusing number of the ircrfts for the uncorrupted pixels)[9] detection scheme is described in this section for use before the medin filtering for the restortion, s result, the proposed restortion method bsed on the restricted medin cn keep the imge unchnged when the filter processing window moves cross the uncorrupted imge detils. Clerly, it my become expensive to perform sort on pixels within lrge rectngulr window. If the width of processing window is lrger thn three, modified medin filter cn be pplied lterntively in the ASCF technique. The modified medin will be bsed in the prt of the pixels which from the out line of the window or the noise- free pixels within the processing window (3), since prt of the pixels inside the window my hve been corrupted. The noisy centrl pixel nd its corrupted eight neighborhood pixels, if detected, will be excluded from the set for the medin filtering. The modified medin filter increses the relibility of the restortion nd reduces the computtionl cost, especilly for removing impulses of high noise rtio. If the chnnel correltion, fctor r, for ech corrupted pixel s defined in (1), pproches I, the modified vector filtering is recommended The number of fuzzy membership functions ssocited with ech vrible depends on the denoising opertions nd the sum of the fuzzy membership vlues where the functions overlp is recommended to be one or less thn one. Becuse 537

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 the AGFF cn tolerte the estimtion devition of the noise rtios, the simple trpezoidl shpe is chosen s the functions in the fuzzifiction process. The mximum method is used in defuzzifiction. The noise type of the slt-nd-pepper cn be determined by the vlues of s 1 or s 2. Noisy restore color imges with medium or high noise rtio. Opertion III Consists of two psses of Opertion I, Opertor D nd one-pss filtering to restore pixels in S 3. First, one-pss filtering of pixels in S 3 is pplied, which is followed by Opertion I. The second pss includes Opertor D which is followed by nother Opertion I. Opertion III is designed to restore corrupted imges with high noise rtio. Fuzzy peer group Impulse detection bsed Impulse Detection Opertion IV Consists of two psses of Opertion III nd one-pss filtering of pixels in S 3 to restore color imges with impulse of very high noise rtio. It pplies Opertion III nd then restores impulse corrupted pixels in S 3 where for in (7). Finlly, it repets Opertion 6. RESULTS OF THE IMPULSE NOISE FILTER. Averge Fuzzy Filtered Imge ASCF Fig 1. Block Digrm Of ASCF Filtering Stge A design principle for the following opertions, which re dpted to different noise rtios nd types, is to use s smll size of the window nd s less number of the psses s possible, s long s the impulse noise cn be removed (to ensure preserving imge detils s much s possible). The number of psses ws determined for removl of noise region bsed on the worst cse scenrio within the estimted mximl size of the noise region. The opertions designed for removing impulses from different corrupted pixel sets in nturl digitl color imges, re defined s follows. Opertion I Consists of two-pss filtering to restore color imges with low noise rtio. In the first pss, it restores impulse corrupted pixels in S 2. In the second pss, it restores impulse corrupted pixels in S 1. Opertion II Consists of Opertion I nd Opertor D. The Opertor D is designed to remove corrupted pixels in S 4. Opertion II is designed to Window sizes of 3 3, 5 5 nd 7 7 re experimented. A plot between PSNR nd percentge of impulse noise for this window size is drwn in Figure 3. The best results for higher percentges of the impulse noise, lrger widow seems to be more pproprite but this filter is less suitble for high level of noise s there is loss of imge detils. As the window size of 3 3 produces better results up to 20% impulse noise, this filter is ment to del with low nd middle percentges of the impulse noise. This level of noise is usully found in mny prcticl pplictions. The performnce of this filter is illustrted through set of color imges with the impulse noise of densities 10%, 15nd 20%. A comprtive nlysis of the proposed techniques is crried out with respect to two recent pproches in the literture, nmely, SMDE method proposed by Pei-Eng Ng et l. [19] nd Luo s EDPA [20]. A smple set of the originl imges used in the experimenttion re displyed in the vlues of MSE nd PSNR enumerted in Tble 1 for different experiments indicte tht the proposed method is ble to reduce more noise from the imges while preserving lmost ll imge detils. The results re better thn those reported in the literture s demonstrted by higher vlue of PSNR in most of the imges nlyses. It cn lso be observed visully tht the proposed filters re quite effective in noise reduction. The results of denoising obtined by few existing methods in the literture re shown in Figure 2 including the results chieved by the proposed impulse filter for 538

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 comprtive purposes. while input to ech filter hve the sme level of noise () Len Imge with Impulse noise of density 15%, (b) Len with FNRC, (c) Len with NRFF, (d) Len with Proposed, (e) Fish Imge with Impulse noise of density 15%, (f) Fish with FNRC, (g) Fish with NRFF, (h) Fish with Proposed, (i) Bird Imge with Impulse noise of density 15%, (j) Bird with FNRC, (k) Bird with NRFF, nd (l) Bird with Proposed method. A plot (See Figure 4) between PSNR nd percentge of Impulse noise for different methods proves this point for the Len imge (i) (j) () (b) (k) (l) Figure 2: Denoised Imges obtined with different Filters (c) (d) Figure 3: PSNR vs. Window sizes for Impulse noise (e) (f) (g) (h) Figure 4: A comprtive nlysis of impulse noise reduction 539

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 7. CONCLUSION A sttisticl complexion-bsed filtering technique hs been proposed for removing impulse noise from corrupted digitl color imges. The specil contribution of the new filtering technique is its novel impulse detection method, which uses 2-D geometric fetures (shpe nd edge type) nd the size of the impulse corrupted pixel/pixel region, insted of 1-D sttisticl informtion, to identify the impulse in n effective nd efficient mnner. The other novelty is its progressive dptive restortion mechnism, where crefully selected set of sizes nd shpes of processing windows re employed, dpting to noise rtio nd type to recover the corrupted pixels step by step through relible multipss process of low computtionl complexity This technique lso provides very relible impulse noise type nd rtio discrimintion method. Through extensive experiment conducted using wide rnge of nturl color imges, the proposed filtering technique hs demonstrted superior performnce to tht of well-known benchmrk techniques, in terms of stndrd objective mesurements, visul imge qulity nd the computtionl complexity, in removing the sltnd-pepper nd the rndom impulse noise which re commonly considered in color imge restortion. The technique is very useful for online pplictions to suppress impulse noise especilly for medium nd lrge sized color imges. It cn be further integrted with other benchmrk techniques to suppress mixed Gussin nd impulse noise contmintion for color imges to improve their performnce REFRENCES: [1] I. Pits nd A. N. Venetsnopoulos, Nonliner Digitl Filter: Principles nd Applictions. Norwell, MA: Kluwer, 1990 [2] E. Abreu, M. Lighstone, S. K. Mitr, nd K. Arkw, A new efficient pproch for the removl of impulse noise from highly corrupted imges, IEEE Trns. Imge Process., vol. 5, no. 6, pp. 1012 1025, Jun. 1996. [3] V. I. Ponomryov, Rel-time 2D-3D filtering using order sttistics bsed lgorithms, J. Rel-Time Imge Process., vol. 1, no. 3, pp.173 194, 2007 [4] X. Li, On modeling interchnnel dependency for color imge denoising, Int. J. Img. Syst. Technol., vol. 1, no. 3, pp. 163 173, Oct. 2007. [5] O. Lezory, A. Elmotz, nd S. Bougleux, Grph regulriztion for color imge processing, Comput. Vis. Imge Understnd., vol. 10, no. 1 2, pp. 38 55, Jul. Aug. 2007. [6] A. Elmotz, O. Lezory, nd S. Bougleux, Nonlocl discrete regulriztion on weighted grphs: A frmework for imge nd mnifold processing, IEEE Trns. Imge Process., vol. 1, no. 7, pp. 1047 1060, Jul. 2008. [7] P. E. Trhnis nd A. N. Venetsnopoulos, Vector direction filter: A new clss of multichnnel imge processing filter, IEEE Trns. Imge Process., vol. 2, no. 10, pp. 528 534, Oct. 1993 [8] D. G. Krkos nd P. E. Trhnis, Combining vector medin nd vector direction filters: The directionl-distnce filter, in Proc. IEEE Int. Conf. Imge Process., Wshington, DC, Oct. 1995, vol. 1, pp. 171 174. [9] R. Lukc nd K. N. Pltniotis, A txonomy of color imge filtering nd enhncement solutions, in Advnces in Imging nd Electron Physics, P. W. Hwkes, Ed. New York: Elsevier, 2006, vol. 140, pp. 187 264. [10] V. Chtzis nd I. Pits, Fuzzy sclr nd vector medin filters bsed on fuzzy distnces, IEEE Trns. Imge Process., vol. 8, no. 5, pp. 731 734, My 1999. [11] H. H. Tsi nd P. T. Yu, Genetic-bsed fuzzy hybrid multichnnel filters for color imge restortion, Fuzzy Sets Syst., vol. 114, pp. 203 224, 2000. [12] R. Lukc, K. N. Pltniotis, B. Smolk, nd A. N. Venetsnopoulos, cdna microrry imge processing using fuzzy vector filtering frmework, J. Fuzzy Sets Syst., vol. 152, no. 1, pp. 17 35, My 2005 [13] L. Khriji nd M. Gbbouj, Adptive fuzzy order sttistic-rtionl hybrid filters in color imge processing, Fuzzy Sets Syst., vol. 128, no. 1, pp. 35 46, Mr. 2000. [14] E. S. Hore, B. Qiu, nd H. R. Wu, Improved color imge vector filtering using fuzzy noise detection, Opt. Eng., vol. 42, no. 6, pp. 1656 1664, Jun. 2003. [15] A. C. Bovic, Streking in medin filtered imges, IEEE Trns. Acoustic., Speech, Signl Process. vol. 35, pp. 493 503, Oct. 1985. [16] B. Smolk, K. N. Pltniotis, nd A. N. Venetsnopoulos, Nonliner techniques for color imge processing, in Nonliner Signl nd Imge Provessing: Theory, Methods, nd 540

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 Applictions. Boc Rton, FL: CRC, Dec. 2003, pp. 445 505 [17] R. C. Gonzlez nd R. E.Woods, Digitl Imge Processing, 2nd ed. Englewood Cliffs, NJ: Prentice-Hll, 2002. [18] Y. Nie nd K. E. Brner, The fuzzy trnsformtion nd its ppliction in imge processing, IEEE Trns. Imge Process., vol. 15, no. 4, pp. 910 927, Apr. 2006 [19] S. Schulte, V. D. Witte, M. Nchtegel, D. V. Weken, nd E. E. Kerre, Fuzzy two-step filter for impulse noise reduction from colour imges, IEEE Trns. Imge Process., vol. 15, no. 11 pp. 3568 3579, Nov.2006. [20] Z. M, H. R. Wu, nd B. Qiu, An structure dptive hybrid vector filter for the restortion of digitl color imges, IEEE Trns. Imge Process., vol. 14, no. 12, pp. 1990 2001, Dec. 2005. [21] Z. M, H. R. Wu, nd D. Feng, Fuzzy vector prtition filtering technique for color imge restortion, Comput. Vis. Imge Understnd., vol. 107, no. 1 2, pp. 26 37, Jul. Aug. 2007. 541

Journl of Theoreticl nd Applied Informtion Technology 31 st Jnury 2014. Vol. 59 No.3 ISSN: 1992-8645 www.jtit.org E-ISSN: 1817-3195 Tble 1: Comprison Of Performnce For Impulse Noise Imge Noisy SMDE EDPA Proposed MSE PSNR MSE PSNR MSE PSNR MSE PSNR Len 10% 11.40 37.20 5.18 40.99 1.85 45.46 1.47 47.46 15% 19.05 34.33 5.94 40.39 2.85 43.58 2.24 45.63 20% 25.28 34.10 7.27 39.52 3.89 42.23 3.12 44.19 Fish 10% 12.78 37.06 10.04 38.11 4.23 41.87 4.02 43.09 15% 19.12 35.32 10.85 36.78 6.53 40.33 5.53 40.70 20% 25.61 34.05 11.66 37.46 7.84 39.19 7.04 39.66 Bird 10% 12.83 37.05 9.39 38.40 3.28 42.97 2.71 44.80 15% 19.45 35.24 10.79 37.80 4.99 41.15 4.26 42.84 20% 25.80 34.02 12.41 37.19 7.11 39.61 5.99 40.36 542