PERFORMANCE EVALUATION OF SPATIAL FILTERS USING FULL-REFERENCE IMAGE QUALITY METRICS

Transcription

1 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. PERFORACE EVALUATIO OF SPATIAL FILTERS USIG FULL-REFERECE IAGE QUALITY ETRICS Palwider Sigh 1 ad Leea Jai 1 I.K.G Pujab Techical Uiversity, Jaladhar, Idia Global Istitute o aagemet ad Emergig Techologies, Amritsar, Idia E-ail: palwider_gdu@yahoo.com ABSTRACT The de-oisig o digital images is crucial preprocessig step beore movig toward image segmetatio, represetatio ad object recogitio. It is importat to id out eicacy o ilter or dieret oise models because ilterig operatio is applicatio orieted task ad perormace varies accordig to type o oise preset i images. A comparative study has made to elucidate the behavior o dieret spatial ilterig techiques uder dieret oise models. I this paper dieret types o oises like Gaussia oise, Speckle oise, Salt & Pepper oise are applied o grayscale stadard image o Lea ad usig spatial ilterig techiques the values o ull reerece based image quality metrics are oud ad compared i tabular ad graphical orm. The outcome o comparative study shows that Lee, Kua ad Aisotropic Diusio Filter worked well or Speckle oise, the Salt ad Pepper oise has sigiicatly reduced usig edia ad AWF, ad the ea ilter ad Wieer ilter works immesely eiciet or reducig Gaussia oise. Keywords: spatial ilter, additive oise, multiplicative oise, image quality metrics. ITRODUCTIO The use o digital images is rapidly icreasig i the ield o educatio, medical diagosis, astroomy, ad mauacturig idustry. There are umerous other areas i which digital images are beig used. The problems associated with digital images are emergig with the icrease i its applicatio. The degradatios i digital images are maily classiied by [1] as give below: Degradatio due to o zero dimesios o picture elemet used i device used or imagig, which is kow as spatial degradatio. Degradatio due to o zero exposure time o photosesitive material used i device used or imagig, which is kow as temporal degradatio. Degradatio due to distortio o image geometry, which is kow as geometrical degradatios. Degradatio due to modiicatio i gray level o image elemet, which is kow as oise. Our mai ocus i this paper will be o degradatio due to modiicatio o gray level i.e., oise. The umber o problems i digital images ca arises due to oise because the oise damages the importat eatures o the image. oise may occur i digital image durig acquisitio, trasmissio ad retrieval process, ad degrades the quality o image. The purpose o de-oisig is to elimiate oise while preservig edges ad other sharp trasitios []. The removal o oise is still a diicult task or researchers because de- oisig algorithms may cause blurrig o edges ad it may elimiate some importat eatures o image also. The algorithm selected or de-oisig depeds upo the ature o the oise preset i the image ad the type o image to be de-oised because the de-oisig algorithm used or ultrasoud images may ot be suitable or satellite images. Similarly the de-oisig algorithm used or Gaussia oise may ot be suitable or speckle oise. The de-oisig is a process o restorig a image ito accurate state ad or this it must be kow by which oise model the image has bee degraded. Whe a degradatio model is kow the iverse process ca be applied to restore a image back ito origial orm. The image degradatio ad restoratio model [] is give as ollows. ( g ( ˆ( Degradatio Restoratio Fuctio (H) ( Figure-1. Image restoratio/degradatio model. 907

2 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Where ( is the origial image, g ( is oisy image corrupted by some kow degradatio ad additive oise ( ad ˆ( restored image which is approximatio o origial image. I this paper degradatio uctio is assumed to be idetity ad ilterig methods will be based o oise preset i the image. The oise preset is assumed to be idepedet so there is o correlatio betwee pixel value ad oise. The spatial ilterig methods ea, edia, Adaptive Weighted edia Filter(AWF), Wieer, Kua, Frost ad Aisotropic Diusio ilter are applied o stadard image o Lea corrupted with Gaussia oise, Speckle oise ad Salt & Pepper oise respectively. The eicacy o dieret spatial ilter has evaluated usig ull reerece image quality metrics Root ea Square Error (RSE), Peak Sigal to oise Ratio (PSR), Structural Cotet (SC), Average Dierece (AD), aximum Dierece (D), ormalized Correlatio Coeiciet (CC), ormalized Absolute Error (AE), Laplacia ea Square Error (LSE) ad Structural Similarity Idex (SSI). OISE ODELS Image oise is uwated sigal preset i captured image. Images are corrupted by oise durig the process o acquisitio, storig ad trasmissio, like speckle oise is iherited i coheret imagig ad occurs durig image acquisitio. Ultrasoud, sythetic aperture radar imagig are examples o coheret imagig [4]. Durig trasmissio the oise maily occurs i aalogue chaels. The digital chaels restrict the itererece o artiacts with sigal over trasmissio chael. The oise visually degrades the quality o image as well as make diicult to study area o iterest. I this sectio, various oise models, their categorizatios ad their probability distributed uctio will be discussed. The oise is geerally categorized as additive or multiplicative [3]. The Gaussia oise is additive i ature. It is represeted as i ollowig equatio (1). g( ( ( (1) The value o each pixel i the image havig Gaussia oise will be sum o the value o pixel i reerece image ad the value o radom Gaussia distributed oise. This type o oise has a bell shaped, ad The Gaussia probability distributio uctio is give as ollow i equatio (). 1 z P ( z) e () Where represets mea itesity o give image, z represets the gray level ad is the stadard deviatio o oise. The probability distributio uctio o Gaussia oise ad Gaussia oise patter geerated i matlab is give i ollowig Figure-. P(z Figure-. Distributio uctio or Gaussia oise. Figure-3. Gaussia oise patter geerated i atlab. The Speckle oise which is multiplicative i ature emerges whe coheret light illumiates the rough surace the the relected waves rom the surace cotais cotributio o may idepedet scatterig areas [4]. The scattered compoets with relative delay o ew wavelegths sums together while propagatig to distat observatio poit ad orm a graular patter ad that patter is termed as speckle. The speckle oise is itrisic artiact o ultrasoud ad sythetic aperture radar imagig [5, 6]. The quality o the image ca be improved by usig speckle reductio ilterig techiques. The speckle oise is multiplicative i ature ad give i ollowig equatio g( ( * m( ( (3) Where ( is the origial image, g ( is degraded image corrupted by additive oise ( ad multiplicative oise m (. Suppose the eect o additive oise such as sesor oise is egligible as compared to multiplicative oise. ( m( (4) z 908

3 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. equatio (1 ) ca be rewritte as g( ( * m( (5) The logarithmic process coverts the give multiplicative equatio ito additive orm. The probability desity uctio or speckle oise is give below. P( z) z 1 e ( 1)!a z a Where z represets the gray level ad variace is a α. The speckle patter ca be o three types [5] a) Fully ormed speckle, b) o radomly distributed with log rage order, ad c) o radomly distributed with short rage order. This classiicatio is accordig to scatterer umber desity ad spatial distributio. The probability desity uctio o Speckle oise ad patter o speckle oise geerated i matlab is give i ollowig Figure. (6) oise is set to 55 ad value o pepper oise is set to 0. The probability desity uctio P(z) or salt &pepper oise is give as ollow: P m or z m P( z) P or z (7) 0 otherwise Where z represets the gray level. Oce the media was the most eiciet ad eective ilter or removig salt ad pepper oise. But the media ilter cause blurrig o edges ad ca itroduce artiacts whe oise level icrease beyod threshold value [8]. I order to avoid these problems alteratives based o media ilter like adaptive or switchig ilters were proposed which we will discuss i ext sectio. The graphical represetatio o probability desity uctio o salt ad pepper oise ad patter o salt ad pepper oise geerated i matlab is give as ollow: P(Z) P(z P(b P(a Figure-4. Distributio uctio or Speckle oise. a b z Figure-6. Distributio uctio or salt ad pepper oise. Figure-5. Gaussia oise patter geerated i atlab. The salt ad pepper oise is basically a impulse oise which is caused by sudde ad sharp disturbaces i camera sesors or trasmissio i oisy chaels [8]. The other type o impulse oise is radom valued oise. The pixel values i image corrupted with salt ad pepper oise are set to either value m or ad the pixels which are uaected remai uchaged. The values m ad are miimum ad maximum value withi dyamic rage. For a 8 bit image havig 56 gray values, the value o salt Figure-7. Gaussia oise patter geerated i atlab. FILTERIG OF DIGITAL IAGES oise may arise i digital images due to umber o actors like improper sesig elemet, improper lightig coditios, or eviromet actors etc. The deoisig o images is very importat pre-processig step i order to get good results i segmetatio, object recogitio ad represetatio. For example, the 909

4 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. ultrasoud images cotai speckle oise durig acquisitio which is udesirable because as it aects the task o huma iterpretatio ad diagosis.but removal o oise rom digital images may itroduce some artiacts ad may also cause blurrig due to which area o iterest i digital image may damage. Alog with de-oisig, it is very importat to preserve edges ad importat characteristics o image. A classiicatio o liear ad o liear ilters ca be made i spatial ilterig [9]. The value o output pixel i liear ilterig is the liear combiatio o the values o the pixels i the eighborhood o iput pixel, whereas output is ot a liear uctio o iput i case o liear ilterig. A Sigiicat work i spatial ilterig has doe by [10] usig mea ilter which is based o local statistical measures. The wieer ilter [11] is cosidered as a best ilter amog the class o liear spatial ilters. The Gaussia oise ca be sigiicatly reduced usig wieer ilter. I [1] Pratt also made a qualitative study o two dimesioal media ilters o dieret size ad shapes. They cocluded that the irrespective o size ad shape o widow the media ilter works well or suppressig salt ad pepper oise. The media ilter belogs to a class o o liear spatial ilters. For reducig speckle oise, L mea ilter [13], Adaptive weighted media ilter [14] ad ilter based o o liear diusio [15] belogs to post ormatio ilterig methods which works directly o origial image were best. I [16] ew adaptive ilterig algorithm was used i which arithmetic ilter was used or homogeous regios ad media ilter was used or edge pixels. I[3] 1990, peroa ad malik proposed a ovel ilterig usig aisotropic diusio based o the ilters proposed by Lee [0], Frost [1] ad Kua [].These ilters are based o miimizig mea square error. The use o o liear PDE methods ivolvig aisotropic diusio has sigiicatly icreased. I [] oise smoothig adapts accordig to chages i o-statioary local mea ad o-statioary local variace. I this paper eicacy o some o the existig ilterig techiques like ea, edia, AWF, Lee, Kua, Frost ad Aisotropic diusio ilter has compared usig ull reerece objective image quality metrics. ea ilter The mea ilter is used to reduce itesity variatio betwee pixels by providig smoothess i the give digital image. The itesity variatio is suppressed but blurrig o edges ad itroductio o artiacts may occur i image iltered by mea ilter. It is a average ilterig techique which meas a mask is applied over each pixel ad pixel value is replaced by average o all eighborig pixels value [11]. Figure-8. A example o ea Filter. ask size begis with 3 ad may icrease up to iite limit but as mask size icreases the blurrig may also icrease. The problem with mea ilter is that sometimes a sigle pixel with extreme value sigiicatly aects its eighbor pixels. The mea ilter is give as ollow: 1 ( m ˆ (8) ( x, y) g( x, y) S i, j Where ˆ( is de-oised image, g ( is degraded image ad S i, j is the area deied by mx size mask. edia ilter edia ilter is the order-statistical ilter which works well or salt ad pepper oise but ot suitable or Gaussia ad speckle oise. The value o pixel ( i iltered image will be media o all values give i eighborhood o ( i origial image. The eighborhood widow may be the size o 3x3, 5x5, 7x7 etc. The size o widow will always be i odd umbers. The widow selected is moved to etire image i the same way to mea ilter. It also preserves the details ad does ot itroduce artiacts i the digital image while de-oisig ulike liear ilters such as mea ilter which may blur the edges [11,1]. The reaso is media ilter is ot much sesitive to extreme values ulike mea ilter. It belogs to the category o o-liear ilters. The media ilter is deied as ollow: ˆ media ( ( x, y) (9) ( x, y) w i, j Where ( is the reerece image, ˆ( is de-oised image, ( is the pixel i ( to be processed ad ( x, y) is local widow. It is easy implemet ad easy to ormulate. The algorithm or media ilter is give as Step 1: Cosider each pixel i the image ad select 3x3 size eighborhood. Step : Sort all the pixels i give eighborhood ito umerical order

5 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Step 3: Fid out media value. Step 4: Replace the give pixel with calculated media value. Step 5: Repeat all above steps util all the pixels i the give image are replaced with media values. Adaptive weighted media ilter The adaptive media ilter works o the priciple that the behavior o ilter applied is chages whe statistical measures like mea, variace chages withi image. It also chages size o eighborhood widow durig executio. The basic priciple o algorithm remai same that we eed to id media with i eighborhood widow but value o give pixel will be replace with media or ot, it depeds upo statistical measures [17]. The algorithm or adaptive media ilter is give as ollow: Step 1: Select iitial widow_size ad maximum_widow_size. Step : Cosider ext pixel o image ad repeat ollowig steps or each ad every pixel o image. Step 3: Fid out i_value, ax_value ad edia_value withi give eighborhood. Step 4: I i_value<edia_value<ax_value I i_value<pixel_value<ax_value the retur Pixel_value ad goto step- else retur edia_value ad goto step- else iwidow_size<maximum_widow_size the icrease size o eighborhood widow adgoto step-4 else retur pixel_value ad goto step- The i_value ad ax_value are miimum ad maximum itesity values, geerally cosidered as compoet o impulse oise. We start with idig edia_value ad i it all betwee ax_value ad i_value ad I pixel value does ot alls betwee ax_value ad i_value the it is cosidered as oise ad get replaced with media value o eighborhood widow, otherwise the pixel value will remai ualtered. but i media value itsel is extreme the widow size is icreased ad algorithm restart rom begiig [17]. It works well or impulse oise but does ot give good results or speckle ad Gaussia oise. Wieer ilter The Wieer Filter belogs to a class o optimum statioary liear ilter to ilter images degraded by additive oise. There is a eed to take assumptio i wieer ilter that sigal ad oise are secod order statioary ad perormace criteria is to id iltered image such that ea square error betwee degraded image ad iltered image is miimum [18]. It ca also be used i requecy domai as well as spatial domai. I requecy domai it is used or de-oisig ad de-blurrig whereas i spatial domai oly de-oisig purpose is solved [19]. The spatial domai is take as a base o wieer ilterig i this paper. It uses both local statistics ad global statistics or deoisig. For a give degraded image g(, the wieer ilter is give as ollow: ˆ( g( mea itesity, Where ˆ( (10) is de-oised image, is local is local variace, ad is variace o reerece image. It is adaptively applied to oisy image which meas it perorm less smoothig or large variace ad more smoothig or more smoothig or small variace. Lee ilter The Lee Filter [0] is based o local statistical measure or de-oisig ad or preservig edges ad ie details. It is a adaptive ilter which meas it perorms ilterig i variace over a area is high otherwise ilterig operatio will igored. It works well or both additive ad multiplicative oise although it was desiged to reduce speckle oise. Iitially mea ad variace o each pixel is derived rom its local mea ad variace the the estimator which miimizes the mea square error is applied to get the de-oisig algorithm. The athematical ormula or Lee ilter is give as ollow: ˆ( W( ) Where W ( g( j 1 g whe W ( 0 whe. We kow that g variace o oisy image ad g g (11) ad is is variace o de-oised image. The rage o W( lies betwee 0 ad 1. Frost ilter I [1] a de-oisig model or radar image was developed or radar images which were corrupted by multiplicative oise. Ultrasoud images ad SAR images are usually corrupted by multiplicative oise ad stadard spatial ilterig techiques do ot work or multiplicative speckle oise. A model or de-oisig was preseted to miimize mea square error or smoothig Radar images. Due to o-coheret behavior the Radar images cotais multiplicative oise ad stadard de-oisig techiques are ot applicable o it. The ilter is applied o spatial domai ad computatioally it is very ecoomical. It is a adaptive wieer ilter which covolves the pixel value withi a ixed size mask. The expoetial impulse respose k is give as: k exp C y ( t (1) 911

6 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Where is ilter parameter, ad t 0 is distace measured rom processed pixel (. The ilter is developed with assumptio that both oise ad sigal are statioary. For radar images, the oise ca be modeled as statioary but o a global basis the radar images are ostatioary [1]. Thereore the ilter is adapted to the chages i local properties. Kua ilter I[] the de-oisig o digital images rom sigal depedet oise is doe by usig ilter or oise smoothig which adapts accordig to chages i ostatioary local mea ad local variace. The ilter behave likes a poit processor while smoothig ucorrelated, sigal depedet oise but or multiplicative oise the ilter behave like lee ilter with some modiicatio which permits various estimators or local variace o image. It is similar to Lee ilter with more accuracy i de-oisig ad edge preservatio because little or o approximatio is required i total derivatio. The ormula or kua ilter is same as that o lee ilter which is give i equatio (11) but value o W( is chaged ad give as ollow: W ( Where 1 1 g (13) is the variace o oisy image ad variace o de-oised image. g Aisotropic diusio ilter Aisotropic Diusio ilter uses the cocept o partial dieretial equatio to reduce oise rom digital image. Idea o modelig a ilter or reducig speckle oise based o diusio ilter were irst proposed by peroa ad malik i [3], they developed model o Speckle reducig aisotropic diusio(srad). It ca reduce speckle oise ad preserve as well as ehace edges usig aisotropic diusio which was ot possible with stadard spatial ilterig methods [4]. It is iterative procedure i which images are cosidered to cosist o sub regios ad diusio ilter works or smootheig withi the regio. It provides otable reiemet i oise reductio ad preservatio o edges whe compared with stadard lee, kua, ad rost ilters. The work later o exteded by weickert i [15] ad proposed the coherece ehacig diusio based o tesor valued diusio ilter or smoothig. The Partial dieretial equatio proposed by peroa ad malik or discrete domai is give as ollow ps i, j tt tt t t t I j I j c I (, p I (, p (14) S Where I t i, j j is the digital image, ( deotes pixel positio i dimesioal grid, t is the step size ad S i, j is is the spatial eighborhood o pixel (. The edge preservatio ad smoothig withi the regio is the mai advatage o aisotropic diusio. The adaptive ilterig approach ad PDE later combied to derive speckle reducig aisotropic diusio. Image quality metrics The quality o image ca be judged o various parameters. The applicatios i which the huma eye is the ultimate observer the subjective quality measures are used but there are some applicatios i which assessmet o image quality should be made automatically o the basis o automated measures without ay huma ivolvemet. The subjective metrics are categorized as sigle stimulus i which subject evaluates the quality o test image without ay reerece or source image ad double stimulus i which test images are evaluated i the presece o source images [6]. The evaluatio ca be made rom sigle to multiple repetitios. The results o subjective metrics ca be iterpreted o ew scales, like average, good, best, below average, poor etc. DSIS DSCQS SDSCE, Subjective Image Quality etrics Siglestimulus Doublestimulus SSR SSCQE Figure-9. The subjective image quality metrics. The objective image quality metrics ca automatically predict the quality o image usig some automated models without ay huma assistace. The aim o image quality is to improve overall appearace o image. oreover it ca also be used or ollowig applicatios [4, 5]. It ca be used to stadardize a image or algorithm or ehacemet, compressio or restoratio. It ca be used to keep check o quality o image. It ca be used to optimize the give algorithm. The objective image quality metrics ca be urther classiied as Full-reerece, Reduced-reerece ad o-reerece. 91

7 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Full- Reerece Figure-10.The objective image quality metrics. I Full reerece approach the quality o test image is compared with reerece image. So reerece image i this techique is assumed to be kow. I some cases the reerece image may ot be available the we ca use o-reerece or Blid method as Image quality metric. Whe reerece image is partially available the Reduced- reerece methods are used [8]. The study i this paper is made o Full-reerece image quality metrics. I this paper we have cosidered ( as a reerece image, ˆ( as a iltered image ad * as a size o image. Root mea square error The mea square error is the sum o the square o dierece betwee de-oised image ad reerece image itesity divided by size o image. A higher value o mea square error meas the dierece betwee de-oised image ad reerece image is large or image has ot bee de-oised [7]. athematically mea square error is deied as ollow: 1 ( ˆ( ( ) 1 i1 j (15) The major drawback o SE is its depedece o itesity scalig. The root mea square ca be simply oud by takig square root o the value o mea square error. Peak Sigal to oise Ratio (PSR) The quality o image is measured as a ratio o maximum itesity available (Power i case o sigal) to the square root o mea square error [7]. It is measured i decibels. For a give image (, the PSR is mathematically deied as ollow: S 10 (16) log SE Objective Image Quality Assessmet Reduced- Reerece o- Reerece Where S is maximum itesity ad SE is the mea square error. A higher value o PSR meas the deoised image is o good quality ad is close to reerece image. ea absolute error It is the average o absolute dierece betwee de-oised image ad reerece image. It is used to measure how close the de-oised image is to reerece image [9]. The AE is deied as ollow. 1 i1 1 j ˆ( ( (17) A lower value o AE meas the de-oised image is close to reerece image. aximum dierece It is used to measure maximum dierece betwee de-oised image ad reerece image. A large value o D meas de-oised image is o poor quality. I simple words it is used to measure the maximum error [9]. The D is deied as ollow: max ˆ( ( (18) Structural cotet It is a ratio betwee sums o square o itesities o reerece image to sum o square o itesities o deoised image [9]. The value o SC will be close to 1 or good quality image. The SC is deied as ollow: i1 i1 j1 j1 ( ( ) ( ˆ( ) (19) ormalized absolute error It is used to id error predictio accuracy o the de-oised image. The SC is deied as ollow [9]. i1 j1 i1 ˆ( j1 ( ( (0) ormalized cross correlatio It is a correlatio uctio which is used to measure closeess betwee reerece image ad de-oised image [9]. It is used to measure similarity, i two images are idetical images the value o K will be 1. The K is deied as ollow 913

8 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. i1 j1 i1 ( ) ˆ( j1 ( ( ) (1) Laplacia mea square error Local cotrast is havig a crucial role i deiitio image quality. The LSE is a method used or evaluatig local cotrast o image [9]. The LSE is deied as ollow Figures 11 ad 1 shows ilterig results o image corrupted usig salt ad pepper oise ad image corrupted usig Gaussia oise. The values o objective quality metrics o oisy ad iltered images are give i Table-1, ad 3. Visual comparisos o objective quality metrics o oisy ad iltered images are give i Figures 13 to 18 i the orm o bar charts. i1 j1 ( L( ˆ( ) L( ( )) i1 j1 L( ( ) () Where L( ( ) i 1, j i 1, j, j 1 j 1 4 ( j ad L( ˆ( ) ˆ i 1, j ˆ i 1, j, ˆ j 1 ˆ j 1 4 ˆ( j ) ) Structural similarity idex It is used to quatiy structural chages which iclude lumiace, cotrast ad texture o digital image. The greater value o SSI meas greater similarity betwee reerece image ad de-oised image [30]. The SSI is deied as ollow: ˆ C1, ˆ C ˆ C ˆ C 1 (3) Figure-11. a is image corrupted with Speckle oise, b is image iltered with ea ilter, c is image iltered with edia ilter, d is image iltered with Adaptive_media ilter, e is image iltered with Wieer_ilter, is image iltered with Lee ilter, g is image iltered with Kua ilter, h is image iltered with Frost ilter, i is image iltered with Aisotropic_Diussio ilter. Where is the mea itesity o reerece image, is mea itesity o de-oised image, ˆ is variace o reerece image, image, image., ˆ ˆ is variace o de-oised is Covariace o reerece ad de-oised RESULTS AD DISCUSSIOS The algorithms are compared usig experimetal evaluatio, by addig speckle oise o variace 0.1, Salt & Pepper oise o variace 0.1 ad Gaussia oise o mea 0 ad variace 0.1 o stadard grayscale test image o Lea o size 51x51 ad class uit8 take rom USC- SIPI image database. The quality o iltered image is compared usig ull reerece objective quality measures PSR, RSE, AD, SC, CC, D, LSE, AE ad SSI. The implemetatio is doe o atlab R016B (Versio 9.1). The oisy image ad iltered images obtaied by applyig various spatial ilterig techiques are give i Figures 10, 11 ad 1 where Figure-10 shows image corrupted with speckle oise ad images iltered usig various spatial ilterig techiques. Similarly Figure-1. a is image corrupted with Salt & Pepper oise, b is image usig ea ilter, c is image usig edia ilter, d is image usig AWF, e is image usig Wieer_ilter, is image usig Lee ilter, g is usigkua ilter, h is image usig Frost ilter, i is usig Aisotropic ilter. 914

9 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Figure-13. a is image corrupted with Gaussia oise, b is image usig ea ilter, c is imageusig edia ilter, d is image usig AWF, e is image usig Wieer_ilter, is imageusig Lee ilter, g is usig Kua ilter, h is image usig Frost ilter, i is usigaisotropic ilter. Table-1.The values o RSE, PSR, AD, SC, CC, D, LSE, AE ad SSI i the presece o Speckle oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. RSE PSR AD SC CC D LSE AE SSI Speckle oise ea edia AWF Wieer Lee Kua Frost Aisotropic

10 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. Table-.The values o RSE, PSR, AD, SC, CC, D, LSE, AE ad SSI i the presece o Salt & Pepper oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. RSE PSR AD SC CC D LSE AE SSI Salt&Pepper oise ea edia AWF Wieer Lee Kua Frost Aisotropic Table-3.The values o RSE, PSR, AD, SC, CC, D, LSE, AE ad SSI i the presece o Gaussia oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. RSE PSR AD SC CC D LSE AE SSI Gaussia oise ea edia AWF Wieer Lee Kua Frost Aisotropic RSE PSR AD D LSE Figure-14. Compariso o RSE, PSR, AD, D ad LSE i the presece o Salt & Pepper oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. 916

11 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved SC CC AE SSI Figure-15. Compariso o SC, CC, AE ad SSI i the presece o Salt & Pepper oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters RSE PSR AD D LSE Figure-16. Compariso o RSE, PSR, AD, D ad LSE i the presece o Speckle oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters SC CC AE SSI Speckle ea edia AWF Wieer Lee Kua Frost Diusio Figure-17. Compariso o SC, CC, AE ad SSI i the presece o Speckle oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. 917

12 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved RSE PSR AD D LSE Figure-18. Compariso o RSE, PSR, AD, D ad LSE i the presece o Gaussia oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters SC CC AE SSI Gaussia ea edia AWF Wieer Lee Kua Frost Diusio Figure-19. Compariso o SC, CC, AE ad SSI i the presece o Gaussia oise ad i images iltered usig ea, edia, AWF, Wieer, Lee, Kua, Frost ad Aisotropic diusio ilters. COCLUSIOS I this paper we have worked o eight dieret spatial ilterig techiques usig ie ull-reerece based image quality metrics. A comparative study o spatial ilterig techiques o stadard test image o Lea corrupted o Gaussia, speckle ad salt ad pepper oise was carried out. We obtaied very useul results which depicts that each ilter works well or certai type o oise models ad does ot work so good or other models. The aalysis was doe usig subjective iterpretatio o iltered images as well as usig ull reerece based image quality metrics. By comparig the results o image quality metrics the coclusio is made that the Speckle oise ca be reduced usig Lee, Kua ad Aisotropic diusio ilter, Salt ad Pepper oise ca be suppressed usig edia ad AWF ad or Gaussia oise mea ad wieer ilter are immesely eiciet. Although we got good results but still there is a scope o improvemet i SSI i the presece o speckle oise. Oe o the mai uture directios is to apply trasorm ilterig i wavelet domai o images corrupted by speckle oise. ACKOWLEDGEET We are grateul to I.K.G Pujab Techical Uiversity, Jaladhar or their support ad ecouragemet. REFERECES [1]. Soka, V. Hlavac ad R. Boyle Image processig, aalysis ad machie visio. Thomso, Toroto: Caada. [] R. C. Gozalez ad R.E. Woods Digital Image Processig. Pearso Pretice Hall, ew Jersey. [3] I. Pitas ad A.. Veetsaopoulos o liear Digital Filters: Priciples ad Applicatios.The Spriger Iteratioal series i Egieerig ad Computer sciece. [4] J. W. Goodma Some udametal properties o speckle. Joural o optical society o America. 66(11):

13 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. [5] C. B. Burckhardt Speckle i Ultrasoud B- mode scas. IEEE trasactio o Soics ad Ultrasoics. 5(1): 1-6. [6] Q. a. ad D. Kapla o the statistical characteristics o log-compressed Rayleigh sigals: theoretical ormulatio ad experimetal results. Joural o acoustical society o America. 3: [7].C. otwa R.C. otwa F.C. Harris ad.c 004.Gadiya Survey o image deoisig techiques. Proceedigs o Global Sigal Processig, Sata Clara. [8] C. Bocelet Image oise odels. Hadbook o image ad Video Processig, Edited by A. C. Bovik. [9] J.W. Tukey o liear methods or smoothig data. Proceedig o EASCO. p [10]. S. Jayat Average ad media based smoothig techiques or improvig digital speech quality i the presece o trasmissio error. IEEE trasactio o commuicatios. 4: [11] A. K. Jai.1989.Fudametals o Digital Image Processig. Pretice Hall iormatio ad system scieces series. [1] W. K. Pratt Digital image processig 4 th editio.wiley, ew York, USA. [13] C. Kotropoulos ad I. Pitas Optimum o liear sigal detectio ad estimatio i the presece o ultrasoic speckle. Ultrasoic Imagig. 14(3): [14]. Karama,. A. Kutay ad G. Bozdagi A adaptive speckle suppressio ilter or medical ultrasoic imagig. IEEE trasactio o medical imagig. 14():83-9. [15] J. Weickert Eiciet ad reliable schemes or o liear diusio ilterig. IEEE trasactio o image processig. 7(3): [16]. Rakovic ad. Tuba. 01. Improved Adaptive edia ilter or deoisig ultrasoud images. Proceedigs o the 6 th Europea computig coerece. pp [17] E. Atama, K.. Wog ad B. K. Aatre Some statistical properties o media ilter. IEEE Trasactio o acoustics, speech ad sigal processig. 9: [18] L. Gua ad R. Ward Restoratio o Radomly blurred images by the Wieer Filter. IEEE Trasactio o acoustics, speech ad sigal processig. 37(4): [19] S. Kumar ad P. Kumar Perormace Compariso o edia ad Wieer Filter i Image Deoisig. Iteratioal joural o Computer Applicatios. 1(4): [0] J. Lee Digital image ehacemet ad oise ilterig usig local statistics. IEEE Trasactio o Patter aalysis ad achie itelligece. : [1] V. S. Frost, J. A. Stiles, J. C. Holtzma ad K.S. Shamugam A model or radar images ad its applicatio to adaptive digital ilterig o multiplicative oise. IEEE Trasactio o Patter aalysis ad achie itelligece. 4: [] D. T. Kua, A. A. sawchuk, P. Chavel ad T.C. Strad Adaptive oise smoothig ilter or images with sigal depedet oise. IEEE Trasactio o Patter aalysis ad achie itelligece. 7: [3] J. alik ad P. Peroa Scale space ad edge detectio usig aisotropic diusio. IEEE trasactio o Patter aalysis ad achie itelligece. 1: [4] Y. Yu ad S. T. Acto. 00. Speckle Reducig Aisotropic Diusio. IEEE trasactio o image processig. 11: [5] Z. Wag, H. R. Sheikh ad A. C. Bovik Objective video quality assessmet. I the hadbook o video databases: Desig ad applicatios. Laboratory o image ad video egieerig, The Uiversity o Texas, Austi, CRC Press, ch. 41: [6] T.. Pappas ad R. J. Sarack Perceptual criteria or image quality evaluatio. Hadbook o Image ad Video Processig, Academic Press, Edited by A. C. Bovik. [7] Z. Wag ad A. C. Bovik ea square error, Love it or leave it. IEEE Sigal Processig magazie, DOI /SP [8] K. H. Thug ad P. Raveedra A Survey o image quality measures. IEEE iteratioal coerece or techical postgraduates. pp

14 VOL. 13, O. 3, FEBRUARY 018 ISS Asia Research Publishig etwork (ARP). All rights reserved. [9] A.. Eskicioglu ad P. S. Fisher Image quality measures ad their perormaces. IEEE trasactio o commuicatio. 43(1): [30] Z. Wag ad A. C. Bovik. 00. A Uiversal image Quality Idex. IEEE Sigal processig letters. 9(3): [31] Z. Wag, A. C. Bovik, H. R. Sheikh ad E. P. Simocelli Image quality assessmet: From error measuremet to structural similarity. IEEE trasactio o image processig. 13(4):