基于D维广义高斯分布与层次聚类相结合的纹理分割研究.(IJIGSP-V8-N3-6)

I.J.Image,GraphicsandSignalProcessing,2016,3,45-54PublishedOnlineMarch2016inMECS()DOI:10.5815/ijigsp.2016.03.06Copyright©2016MECSI.J.Image,GraphicsandSignalProcessing,2016,3,45-54StudiesonTextureSegmentationUsingD-DimensionalGeneralizedGaussianDistributionintegratedwithHierarchicalClusteringK.NaveenKumarDepartmentofIT,GIT,GITAMUniversity,VisakhapatnamEmail:[e-mail],[e-mail]Statistics,AndhraUniversity,VisakhapatnamEmail:[e-mail],GIT,GITAMUniversity,VisakhapatnamEmail:[e-mail]fCSE,JNTUK-Kakinada,KakinadaEmail:[e-mail]—Texturedealswiththevisualpropertiesofanimage.Textureanalysisplaysadominantroleforimagesegmentation.Intexturesegmentation,modelbasedmethodsaresuperiortomodelfreemethodswithrespecttosegmentationmethods.ThispaperaddressestheapplicationofmultivariategeneralizedGaussianmixtureprobabilitymodelforsegmentingthetextureofanimageintegratingwithhierarchicalclustering.HerethefeaturevectorassociatedwiththetextureisderivedthroughDCTcoefficientsoftheimageblocks.ThemodelparametersareestimatedusingEMalgorithm.Theinitializationofmodelparametersisdonethroughhierarchicalclusteringalgorithmandmomentmethodofestimation.ThetexturesegmentationalgorithmisdevelopedusingcomponentmaximumlikelihoodunderBayesianframe.TheperformanceoftheproposedalgorithmiscarriedthroughexperimentationonfiveimagetexturesselectedrandomlyfromtheBrodatztexturedatabase.ThetexturesegmentationperformancemeasuressuchasGCE,PRIandVOIhaverevealedthatthismethodoutperformovertheexistingmethodsoftexturesegmentationusingGaussianmixturemodel.Thisisalsosupportedbycomputingconfusionmatrix,accuracy,specificity,sensitivityandF-measure.IndexTerms—MultivariategeneralisedGaussianmixturemodel,texturesegmentation,EM-algorithm,DCTcoefficients,segmentationqualitymetrics.I.INTRODUCTIONThearrangementofconstituentparticlesofamaterialisreferredasTexture.Thetextureisinfluencedbyspatialinterrelationshipsbetweenthepixelsinanimage.Textureusuallyrefersthepatterninanimagewhichincludescoarseness,complexity,fineness,shape,directionalityandstrength[1,2].Severalsegmentationmethodsfortexturesegmentationhavebeendevelopedforanalysingtheimagesconsideringthetexturesurfaces[3-8].Amongthesemethods,modelbasedmethodsusingprobabilitydistributiongainedlotofimportance.Thesemethodsareoftenknownasparametrictextureclassificationmethods.SeveralmodelbasedtextureclassificationmethodshavebeendevelopedusingGaussiandistributionorGaussianmixturedistribution[9-11].ThemajordrawbackofthetexturesegmentationmethodbasedonGaussianmodelorGaussianmixturemodelsisthesegmentationqualitymetricsstillremaininferiortothestandardvaluessuchasPRIclosetoone,GCEclosetozeroandVOIbeinglow.Thisisduetothefactthatthefeaturevectorassociatedwithtextureoftheimageregionsmaynotbemeso-kurtic.Toimprovetheefficiencyofthetexturesegmentationalgorithm,onehastoconsiderthegeneralizationoftheGaussiandistributionforcharacterisingthefeaturevectorassociatedwiththetextureoftheimageregion.Withthismotivation,atexturesegmentationalgorithmisdevelopedandanalysedusingmultivariategeneralizedGaussianmixturemodel.ThegeneralizedGaussiandistributionincludesseveraloftheplaty-kurtic,lepto-kurticandmeso-kurticdistributions.ThisalsoincludesGaussiandistributionasaparticularcase[12].ThefeaturevectorofthetextureassociatedwiththeimageisextractedthroughDCTcoefficientsusingtheheuristicargumentsofYu-LenHuang(2005)[13].Assumingthatthefeaturevectorofthewholeimageischaracterisedby46StudiesonTextureSegmentationUsingD-DimensionalGeneralizedGaussianDistributionintegratedwithHierarchicalClusteringCopyright©2016MECSI.J.Image,GraphicsandSignalProcessing,2016,3,45-54themultivariategeneralizedGaussianmixturemodelwiththefeaturevector,thesegmentationalgorithmbyintegratingheuristicmethodofsegmentation,hierarchicalclustering[14]isdeveloped.Therestofthepaperisorganisedasfollows.Section2dealswiththegeneralizedGaussianmixturemodelanditsproperties.Section3dealswithextractionofthefeaturevectorusingDCTcoefficients.Section4dealswithextractionofmodelparametersusingEMalgorithm.Section5isconcernedwithinitialisationofparameterswithhierarchicalclusteringandmomentmethodofestimation.Section6dealswithtexturesegmentationunderBayesianusingcomponentmixturemodel.Section7dealswithperformanceevaluationofproposedalgorithmthroughexperimentationonfiveimagestakenfromBrodatztexturedataset[15].Section8dealswithcomparativestudyofproposedalgorithmwiththatofothermodelbasedGaussiansegmentationalgorithms.Section9istopresenttheconclusionsalongwithfuturescopeforfurtherresearchinthisarea.II.MULTIVARIATEGENERALIZEDGAUSSIANMIXTUREMODELIntextureanalysis,theentireimagetextureisconsideredasaunionofseveralrepetitivepatterns.Inthissection,webrieflydiscusstheprobabilitydistribution(model)usedforcharacterizingthefeaturevectorofthetexture.Afterextractingthefeaturevectorofeachindividualtextureitcanbemodeledbyasuitableprobabilitydistributionsuchthatthecharacteris