Split and Merge Based Quantitative Approach to Select Filter Combination for Image Segmentation

Transcription

1 0100 (2016) DOI: / mateccof/ Split ad Merge Based Quatitative Approach to Select Filter Combiatio for Image Segmetatio Amit Goyal 1 ad Mohit Srivastava 2 1 Research Scholar, Departmet of ECE, Jaga Nath Uiversity, Jaipur, Rajastha, Idia 2 Professor, Departmet of ECE, Chadigarh Egieerig College, Ladra, Pujab, Idia Abstract. With the advet of image aalysis ad computatio i differet domais, image segmetatio has emerged as the most crucial step to achieve a compact segmet-based descriptio of image scee by decomposig it ito meaigful segmets of similar attributes. The pre-ad-post filterig operatio reduces the effect of oise from the segmeted image. The Camerama image is pre-filtered usig Laplacia, ad filter. The Split ad Merge method for Regio based image segmetatio which guaratees to coected regios are ow applied o the filtered image. The, Laplacia ad Sobel filter is the used to post-filter the segmeted image. The PSNR ad MSE values are calculated to quatitative evaluatio of segmeted images. The quatitative evaluatio of post-filtered segmeted image shows that media filter produces most effective result with lowest MSE of db ad highest PSNR of 5.72 db. 1 Itroductio To aalyze or iterpret a image automatically, preprocessig is doe which ivolves segmetig the image ito differet objects of iterest e.g. separatio of foregroud from the backgroud [6]. Image segmetatio has firmed its groud i may practical applicatios that ivolve a visual iterpretatio, amely i medical imagig, object detectio (face detectio, pedestria detectio, brake light detectio etc. More accurate the segmetatio, more successful is the recogitio of objects i the image. Basically, segmetatio is the process of subdividig a image ito its costituet regios or objects. Segmetatio algorithms are based o either similarity or discotiuity of itesity/gray level values. A umber of image segmetatio algorithms with icreased complexity have bee developed over the years. All these algorithms work o the use of ay of the three mai criteria: the homogeeity withi a segmet, separatio from adjacet segmets ad shape homogeeity. Typically, the segmetatio algorithms ca be grouped ito three major categories o the basis of their segmet formatio properties, amely Threshold Based Segmetatio [11, 22], Boudary based Segmetatio [9, 16] ad Regio Based Segmetatio [10]. 1.1 Features of various segmetatio techiques Threshold based segmetatio produces regios of uiformity withi the give image based o some threshold criteria o the attributes of pixels [11].These algorithms are computatioally less expesive but causes oise, blurred edges or outlier [8] i the image. Cotrary to this, the boudary based techiques rely o edges foud i a image ad tries to locate poits of 1 Correspodig author: amitgoyal0810@gmail.com discotiuities i images. This segmetatio method is very sesitive to oise ad hece post-processig is required. However, regio based segmetatio algorithms partitio the etire image ito sub regios depedig o some homogeeity criteria [, 14]. 1.2 Regio Split ad Merge Techique The regio split ad merge techique falls uder regio based segmetatio ad is combiatio of top-dow ad bottom-up approach [4]. I this image is first split ito homogeeous regios o the basis of homogeeity criteria. The regios formed after first split will be idividually checked for the homogeeity ad ay two adjacet homogeeous regios will be merged ito oe regio. The mergig of homogeeous coected regios is a combiatio of two phases: first is to fid the pair of adjacet homogeeous regios ad secod is to choose a certai mergig criterio [2, 4, 10]. The algorithm halts whe o further mergig is possible. This techique produces large umber of regio boudaries, other tha the horizotal ad vertical oes [1]. But the algorithm of this approach leads to icrease i computatioal speed [10] with lesser complexities ad hece improves segmetatio quality [4]. It also guaratees coected regios [9] i.e. there are o gaps due to missig edge pixels. Due to the iheret advatages of regio based segmetatio techique, the performace of image segmetatio is elaborated i this paper usig split ad merges techique. 2 Methodology Adopted To apply pre ad post filter techiques i order to segmet the image usig split ad merge techique ad evaluate the performace of algorithm, the system architecture beig followed is explaied with the help of flowchart as give i fig. 1. The Authors, published by EDP Scieces. This is a ope access article distributed uder the terms of the Creative Commos Attributio Licese 4.0 (

2 0100 (2016) DOI: / mateccof/ Split Image ito Regios Splittig is a recursive process. Quad tree decompositio method is exploited i the curret work to split a image ito four equal sized-quadrats ad it depeds upo a homogeeity criterio. The curret work uses the attribute of itesity of pixels as the homogeeity criterio. The regios formed after first split will be idividually checked for the homogeeity. The equatio (1) deotes the homogeeity criteria for splittig ad mergig the image. This is iterpreted as a regio is split ito four equal sub-regios or quad-regios if the value of Q comes out to be False. Figure 1. Flowchart of methodology for image segmetatio usig split ad merge techique 2.1 Image Database Gray-scale images of differet properties like dimesios (width*height), resolutio (dots per iches) ad size of image (bytes) for storage may be used as the iput dataset. A gray-scale camerama image (fig.2 (a)) of size 512x512 pixels is chose to draw the performace evaluatio of the algorithm used i this work. Fig. 2(b) helps to visualise the pixel detail of a small sectio of boudary differetiatig the object (perso) from backgroud i the image. Figure 2. (a)iput Image to be segmeted (b) Pixel detail of object boudary 2.2 Pre-filterig Pre-filterig is doe to preserve ad sharpe the edge iformatio [5, 15]. Spatial filters offer more versatility as they ca be used for o-liear filterig applicatios as well [7]. The fudametal of filterig a image lies i its covolutio with the mask of pre-defied filter [20]. Sharpeig liear filters (laplacia, usharp ad laplacia of Gaussia filter) ehace the pixel eighbourhood i the image. Cotrary to liear filters, oliear spatial filterig (media, max ad mi filters) is based o oliear operatios with pixels ecompassed by the filter [21]. I equatio (1), Q is called as predicate, m ad σ are the mea (average itesity) ad stadard deviatios of pixels i a quad-regio, respectively, while a ad b are costats whose values are set durig experimetatio. So the average itesity (m) of a particular regio should be greater tha 0 ad below a certai threshold value. The experimetatios i curret work were doe by takig various values of a ad b, but a appropriate homogeeity criterio was obtaied by takig a ad b as 5 ad 200, respectively. 2.4 Merge Coected Homogeeous Regios The output of the splittig stage is multiple regios which may or may ot be homogeeous depedig o whether they fulfil the homogeeity criterio or ot, respectively [12]. The o-homogeeous regios are further subject to splittig ad the homogeeous regios are merged together. The homogeeity criterio based o average itesity as specified by equatio (1) will be followed [2- ]. The ew regio so formed will be cotiuously merged with its adjacet regios util the homogeeity criterio will be satisfied. 2.5 Post-Filterig To make the quad-tree structure based Split ad Merge image segmetatio capable of adaptig to edges of objects i the image, spatial filters may be applied o the segmeted image [, 5]. It ot oly removes oise but may also be helpful to smooth the arbitrary positio ad orietatio of edges of image objects. Preservatio of edges ca also be doe usig liear filters like Sobel which returs edges at those poits where the gradiet of image is maximum [9, 19]. Other alteratives iclude Prewitts method, Roberts method ad Laplacia of Gaussia etc. 2.6 Quality Assessmet The ext step is to evaluate effectiveess of the segmetatio methods by performig quality assessmet through subjective evaluatio as well as quatitative approaches [2]. The subjective evaluatio is based o visual iterpretatio, the objective evaluatio ca be doe usig a umber of widely-used quality metrics. A few of these metrics are: PSNR- the peak sigal-to-oise ratio (db), the mea square error (MSE) which is the squared orm of the differece betwee the origial image ad 2

3 0100 (2016) DOI: / mateccof/ the segmeted image divided by the umber of elemets [17, 18]. Experimetatio ad Results The split ad merge algorithm was implemeted i MATLAB 7.12 ad tested o stadard camerama test image. At first, the image was filtered by applyig differet spatial filters. The parameters of filters like size, paddig optios were varied to obtai better results of filterig. Table 1 depicts the detail of filters alog with parameter variatios. A Laplacia filter was applied o the origial image ad value of alpha was varied from 0 ad 1 for fie tuig. Sice the Laplacia is a derivative operator, it sharpes the image but drives the costat areas to zero. So the origial image is added back to restore the graylevel [7]. The ehacemet i edges of foregroud (camera ma) is clearly visible fig. (a). The performace of LOG (Laplacia of Gaussia filter) teds towards blurrig the image as the stadard deviatio is lower below to 0.5. Variatios i mask size were also made to aalyse the filter. The result of filterig with LOG filter is show i fig. (b). (c) Image filtered with media filter(size ) A filter offers excellet oise reductio capabilities with cosiderably less blurrig (fig. (c)). The usharp filter acts as sharpeig operator ad ehaces edges via subtractig a smoothed versio of a image from the origial image (fig. 4(a)). But it was observed that blurrig was predomiat with icrease i filter size. Other spatial filters amely mi ad max filter were also applied o the test image i fig. 4 (b) ad 4 (c) respectively. Best filtered image as listed i table 2 from each category has bee selected for further processig. S.o. Table 1. Detail of filters applied o origial image Filter 1 Laplacia 2 LOG(Laplacia of Gaussia) Usharp Max Size (Rows x Parameters Colums) 0 Alpha= x2 Stadard 0. 5x5 0. Deviatio= 5x x 1 Order= 1 4x4 Paddig= Zeros 5x5 Zeros 0.2 Alpha= 0.8 2x2 1 Order= 2x 1 Figure 4. (a) Image filtered with Usharp filter() (b) Image filtered with filter (size=2x Figure. (a) Image filtered with Laplacia filter() (b) Image filtered with LOG filter(size 5x5 & σ=0.5 (c) Image filtered with Max filter (size=2x2) Table 2. Filters with better respose o basis of visual iterpretatio S.No. Filter Size Parameter 1 Laplacia 2 Log 5x5 Stadard Deviatio=0.5 Usharp 4 Paddig= 5 2x Order=1 6 Order=1 Split ad merge segmetatio algorithm was applied o these selected filtered images. It is observed that image filtered with ad filter provides a optimal segmetatio as the foregroud ad backgroud were

4 0100 (2016) DOI: / mateccof/ clearly differetiated without losig edge details. LOG ad Usharp filter causes over-segmetatio ad hece those resposes are rejected for ay further process. Best segmeted image obtaied after applyig split ad merge segmetatio techique o filtered image are show i fig. 5 & fig. 6. Post-filterig has bee ow applied o segmeted image to further preserve edges. Laplacia, ad Sobel filters are used for post filterig operatio. The filters which yielded better results are listed i table. The quality of fial segmeted image after post filterig was assessed quatitatively by calculatig the Mea Square Error (MSE) ad Peak Sigal-to-Noise Ratio (PSNR) values of origial image with respect to segmeted image followed by post-filterig. Calculated values of PSNR ad MSE are give i table. Quatitative evaluatio of post-filterig operatio shows that amog these the Laplacia filtratios have low PSNR ad high MSE, thus ot suitable. However, as per quatitative evaluatio the output quality of filter is highest with PSNR (db) of 5.72, 5.68 ad 5.72, ad MSE (db) of 84.89, ad applied over Laplacia, ad filtered image respectively. The filtered images are show i fig. 7. The quatitative evaluatio shows the filtered output quality of Sobel filter is comparatively poor tha with of PSNR (db) of, ad MSE (db) of i all case. Fig. 8 shows the post filtered images usig Sobel filter. These two postfiltered images are iterpreted visually. Table. Quality assessmet of Split ad Merge algorithm through Quatitative approach Pre-Filter Post-Filter Evaluatio MSE PSN Paramet Name Size Parameter Name Size (db) R er (db) Lapla cia Laplac ia Lapla cia 2x 2x 2x Paddig Laplacia Paddi g= Symme tric Alpha= Order= Paddig 6 Order= Edge Search Filter directio = Paddig usig horizotal ad Order=1 Sobel vertical Figure 5. Segmeted image pre-filtered with (a) Laplacia filter (b) LOG filter (c) Usharp filter Figure 6. Segmeted image pre-filtered with (a) filter (b) filter (size=2x) (c) filter (size=) Figure 7. Segmeted image with as post filter ad Prefilter (a) Laplacia (b) c) 4

5 0100 (2016) DOI: / mateccof/ Figure 8. Segmeted image with Sobel as post filter ad Prefilter (a) Laplacia (b) c ) 4 Coclusios The segmeted image after post filterig was assessed subjectively (visual iterpretatio) as well as with the help of quatitative metrics, amely MSE & PSNR. Visual iterpretatio ifers better result of segmetatio with the use of ad Sobel as post filters. However, it ca be deduced from the quatitative aalysis that i case of as post filter the mea square error is less as compared to other post filters ad higher peak sigal-to-oise ratio ca also be obtaied through it ad specifically its combiatio with Laplacia ad as pre-filters yield most effective. Further, performace of Sobel is also acceptable. It is worthwhile to metio that this combiatio is able to segmet the objects i the image while preservig the edges i.e. edges remai coected, hece this combiatio also provides critically segmeted camerama image. Refereces 1. S.L. Horowitz, T. Pavlidis, Picture segmetatio by a tree traversal algorithm, Joural of the Associatio for Computig Machiery, 2, (1976) 2. R. Huki, L. HaegSuk, A Noble Image Segmetatio usig Local Area Splittig ad Mergig Method based o Itesity Chage, Iteratioal Joural of Software Egieerig ad its Applicatios, 7, (201). X. Wu, Adaptive Split-ad-Merge Segmetatio Based o Piecewise Least-Square Approximatio, IEEE Trasactio o Patter Aalysis ad Machie Itelligece, 15, (199) 4. A.B.M. Faruquzzama, N.R. Paiker, J. Arafat, M.A. Ali, G. Sorwar, Literature o Image Segmetatios Based o Split- ad- Merge Techiques, Iteratioal Coferece o Iformatio Techology ad Applicatios, pp Cairs, Australia (2008). 5. H.S. Yag, S.U. Lee, Split ad Merge Segmetatio Employig Thresholdig Techique, IEEE Iteratioal Coferece o Image Processig, pp Sata Barbara (1997) 6. S. Aadurai, R. Shamugalakshmi, Fudametals of Digital Image Processig, Pearso Educatio, Idia (2007) 7. R.C. Gozalez, R.E. Woods, Digital Image Processig. Pearso Educatio, Idia (2009) 8. R. Yogamagalam, B. Karthikeya, Segmetatio Techiques Compariso i Image Processig, Iteratioal Joural of Egieerig ad Techology, 5, (201) 9. P. Thakare, A Study of Image Segmetatio ad Edge Detectio Techiques, Iteratioal Joural of Computer Sciece ad Egieerig,, (2011) 10. D. Kelkar, S. Gupta, Improved Quadtree Method for Split Merge Image Segmetatio, IEEE Iteratioal Coferece o Emergig Treds i Egieerig ad Techology, pp Nagpur, Maharashtra (2008) 11. R. Kadwal, A. Kumar, S. Bhargava, Review: Existig Image Segmetatio Techiques, Iteratioal Joural of Advaced Research i Computer Sciece ad Software Egieerig, 4, (2014) 12. B. Peg, L. Zhag, D. Zhag, Automatic Image Segmetatio by Dyamic Regio Mergig, IEEE Trasactio o Image Processig, 20, (2011) 1. K. Khoshelham, Z. Li, B. Kig, A Split-ad-Merge Techique for Automated Recostructio of Roof Plaes, Photogrammetric Egieerig & Remote Sesig Joural, 71, (2005) 14. A.B.M. Faruquzzama, N.R. Paiker, J. Arafat, Z. Karim, M.A. Ali, Object Segmetatio Based o Split ad Merge Algorithm, IEEE Iteratioal Coferece o Acoustics, Speech ad Sigal Processig, pp Toulouse, Frace (2006) 15. S.M. Ramesh, B. Gomathy, T.V.P. Sudararaja, Detectio of Defects o Steel Surface for Usig Image Segmetatio Techiques, Iteratioal Joural of Sigal Processig, Image Processig ad Patter Recogitio, 7, 2-2 (2014) 16. P.L. Palmer, H. Dabis, J. Kittler, A performace measure for boudary detectio algorithms, Elsevier Computer Visio Image Uderstadig, 6, (1996) 17. M. Sharma, V. Chouha, Objective Evaluatio Parameters of Image Segmetatio Algorithms, Iteratioal Joural of Egieerig ad Advaced Techology, 2, (2012) 18. H. Zhag, J.E. Fritts, S.A. Goldma, Image segmetatio evaluatio: A survey of usupervised methods, Sciece Direct Computer Visio ad Image Uderstadig, 110, (2008) 19. J. Cay, A Computatioal Approach to Edge Detectio, IEEE Trasactio o Patter Aalysis ad Machie Itelligece, 8, (1986) 20. J.S. Lee, Digital image ehacemet ad oise filterig by use of local statistics, IEEE Trasactio o Patter Aalysis ad Machie Itelligece, 2, (1980) 21. A.V. Lugt, Sigal detectio by complex spatial filterig, IEEE Trasactios o Iformatio Theory, 10, (1964) 22. C.M. Pu, N.Y. A, M. Cheg, A Regio-Based Image segmetatio by Watershed Partitio ad DCT Eergy Compactio, IEEE Iteratioal Coferece o Computer Graphics, pp Chia (2011) 2. Y.J. Zhag, A survey o evaluatio methods for image segmetatio, Elsevier Joural o Patter Recogitio, 29, (1999) 5