毕业论文图像处理英文翻译原文

1382IEEETRANSACTIONSONSYSTEMS,MAN,ANDCYBERNETICS—PARTB:CYBERNETICS,VOL.37,NO.5,OCTOBER2007ColorImageSegmentationBasedonMeanShiftandNormalizedCutsWenbingTao,HaiJin,SeniorMember,IEEE,andYiminZhang,SeniorMember,IEEEAbstract—Inthiscorrespondence,wedevelopanovelapproachthatprovideseffectiveandrobustsegmentationofcolorimages.Byincorporatingtheadvantagesofthemeanshift(MS)segmentationandthenormalizedcut(Ncut)partitioningmethods,theproposedmethodrequireslowcomputationalcomplexityandisthereforeveryfeasibleforreal-timeimagesegmentationprocessing.ItpreprocessesanimagebyusingtheMSalgorithmtoformsegmentedregionsthatpreservethedesirablediscontinuitycharacteristicsoftheimage.Thesegmentedregionsarethenrepresentedbyusingthegraphstructures,andtheNcutmethodisappliedtoperformgloballyoptimizedclustering.Becausethenumberofthesegmentedregionsismuchsmallerthanthatoftheimagepixels,theproposedmethodallowsalow-dimensionalimageclusteringwithsignificantreductionofthecomplexitycomparedtoconventionalgraphpartitioningmethodsthataredirectlyappliedtotheimagepixels.Inaddition,theimageclusteringusingthesegmentedregions,insteadoftheimagepixels,alsoreducesthesensitivitytonoiseandresultsinenhancedimagesegmentationperformance.Furthermore,toavoidsomeinappropriatepartitioningwhenconsideringeveryregionasonlyonegraphnode,wedevelopanimprovedsegmentationstrategyusingmultiplechildnodesforeachregion.Thesuperiorityoftheproposedmethodisexaminedanddemonstratedthroughalargenumberofexperimentsusingcolornaturalsceneimages.IndexTerms—Colorimagesegmentation,graphpartitioning,meanshift(MS),normalizedcut(Ncut).I.INTRODUCTIONImagesegmentationisaprocessofdividinganimageintodifferentregionssuchthateachregionisnearlyhomogeneous,whereastheunionofanytworegionsisnot.Itservesasakeyinimageanalysisandpatternrecognitionandisafundamentalsteptowardlow-levelvision,whichissignificantforobjectrecognitionandtracking,imageretrieval,facedetection,andothercomputer-vision-relatedapplications[1].Colorimagescarrymuchmoreinformationthangray-levelones[24].Inmanypatternrecognitionandcomputervisionapplications,thecolorinformationcanbeusedtoenhancetheimageanalysisprocessandimprovesegmentationresultscomparedtogray-scale-basedapproaches.Asaresult,greateffortshavebeenmadeinrecentyearstoinvestigatesegmentationofcolorimagesduetodemandingneeds.Existingimagesegmentationalgorithmscanbegenerallyclassifiedintothreemajorcategories,i.e.,feature-space-basedclustering,spatialsegmentation,andgraph-basedapproaches.Feature-space-basedclusteringapproaches[12],[13]capturetheglobalcharacteristicsoftheimagethroughtheselectionandcalculationoftheimagefeatures,whichareusuallybasedonthecolorortexture.Byusingaspecificdistancemeasurethatignoresthespatialinformation,thefeatureManuscriptreceivedAugust3,2006;revisedDecember10,2006.ThisworkwassupportedbytheNationalNaturalScienceFoundationofChinaunderGrant60603024.ThispaperwasrecommendedbyAssociateEditorI.Bloch.W.TaoandH.JinarewiththeClusterandGridComputingLaboratory,SchoolofComputerScienceandTechnology,HuazhongUniversityofScienceandTechnology,Wuhan430074,China,andalsowiththeServiceComputingTechnologyandSystemLaboratory,SchoolofComputerScienceandTechnology,HuazhongUniversityofScienceandTechnology,Wuhan430074,China(e-mail:[e-mail];[e-mail]).Y.ZhangiswiththeCenterforAdvancedCommunications,VillanovaUniversity,Villanova,PA19085USA(e-mail:[e-mail]).Colorversionsofoneormoreofthefiguresinthispaperareavailableonlineat[7].Althoughthedataclusteringapproachesareefficientinfindingsalientimagefeatures,theyhavesomeseriousdrawbacksaswell.Thespatialstructureandthedetailededgeinformationofanimagearenotpreserved,andpixelsfromdisconnectedregionsoftheimagemaybegroupedtogetheriftheirfeaturespacesoverlap.Giventheimportanceofedgeinformation,aswellastheneedtopreservethespatialrelationshipbetweenthepixelsontheimageplane,thereisarecenttendencytohandleimagesinthespatialdomain[11],[28].Thespatialsegmentationmethodisalsoreferredtoasregion-basedwhenitisbasedonregionentities.Thewatershedalgorithm[19]isanextensivelyusedtechniqueforthispurpose.However,itmayundesirablyproduceaverylargenumberofsmallbutquasi-homogenousregions.Therefore,somemergingalgorithmshouldbeappliedtotheseregions[20],[28].Graph-basedapproachescanberegardedasimageperceptualgroupingandorganizationmethodsbasedonthefusionofthefeatureandspatialinformation.Insuchapproaches,visualgroupisbasedonseveralkeyfactorssuchassimilarity,proximity,andcontinuation[3],[5],[21],[25].Thecommonthemeunderlyingtheseapproachesistheformationofaweightedgraph,whereeachvertexcorrespondstonimagepixeloraregion,andtheweightofeachedgeconnectingtwopixelsortworegionsrepresentsthelikelihoodthattheybelongtothesamesegment.Theweightsareusuallyrelatedtothecolorandtexturefeatures,aswellasthespatialcharacteristicofthecorrespondingpixelsorregions.Agraphispartitionedintomultiplec