(15)Segmentation and Classification of Hyperspectral Images

SegmentationandClassi cationofHyperspectralImagesUsingMinimumSpanningForestGrown

FromAutomaticallySelectedMarkers

YuliyaTarabalka,StudentMember,IEEE,JocelynChanussot,SeniorMember,IEEE,and

JónAtliBenediktsson,Fellow,IEEE

Abstract—Anewmethodforsegmentationandclassi cationofhyperspectralimagesisproposed.Themethodisbasedontheconstructionofaminimumspanningforest(MSF)fromregionmarkers.Markersarede nedautomaticallyfromclassi cationresults.Forthispurpose,pixelwiseclassi cationisperformed,andthemostreliableclassi edpixelsarechosenasmarkers.Eachclassi cation-derivedmarkerisassociatedwithaclasslabel.EachtreeintheMSFgrownfromamarkerformsaregioninthesegmentationmap.Byassigningaclassofeachmarkertoallthepixelswithintheregiongrownfromthismarker,aspectral-spatialclassi cationmapisobtained.Furthermore,theclassi cationmapisre nedusingtheresultsofapixelwiseclassi cationandama-jorityvotingwithinthespatiallyconnectedregions.Experimentalresultsarepresentedforthreehyperspectralairborneimages.TheuseofdifferentdissimilaritymeasuresfortheconstructionoftheMSFisinvestigated.Theproposedschemeimprovesclassi- cationaccuracies,whencomparedtopreviouslyproposedclas-si cationtechniques,andprovidesaccuratesegmentationandclassi cationmaps.

IndexTerms—Classi cation,hyperspectralimages,markerse-lection,minimumspanningforest(MSF),segmentation.

I.INTRODUCTION

MAGECLASSIFICATION,whichcanbede nedasiden-ti cationofobjectsinascenecapturedbyavisionsystem,isoneoftheimportanttasksofaroboticsystem.Ontheoneside,theprocedureofaccurateobjectidenti cationisknowntobemoredif cultforcomputersthanforpeople[1].Ontheotherside,recentlydevelopedimageacquisitionsystems(forinstance,radar,lidar,andhyperspectralimagingtechnologies)capturemoredatafromtheimagescenethanahumanvisionsystem.Therefore,ef cientprocessingsystemsmustbedevel-opedinordertousethesedataforaccurateimageclassi cation.

I

ManuscriptreceivedMay29,2009;revisedSeptember4,2009.ThisworkwassupportedinpartbytheMarieCurieResearchTrainingNetwork“Hyper-I-Net.”ThispaperwasrecommendedbyAssociateEditorD.Goldgof.

Y.TarabalkaiswiththeGrenobleImagesSpeechSignalsandAu-tomaticsLaboratory(GIPSALab),GrenobleInstituteofTechnology(INPG),38402Grenoble,France,andtheFacultyofElectricalandCom-puterEngineering,UniversityofIceland,107Reykjavik,Iceland(e-mail:yuliya.tarabalka@hyperinet.eu).

J.ChanussotiswiththeGrenobleImagesSpeechSignalsandAutomaticsLaboratory(GIPSALab),GrenobleInstituteofTechnology(INPG),38402Grenoble,France(e-mail:jocelyn.chanussot@gipsa-lab.grenoble-inp.fr).

J.A.BenediktssoniswiththeFacultyofElectricalandComputerEngineer-ing,UniversityofIceland,107Reykjavik,Iceland(e-mail:benedikt@hi.is).Colorversionsofoneormoreofthe guresinthispaperareavailableonlineathttp://www.77cn.com.cn.

DigitalObjectIdenti er

10.1109/TSMCB.2009.2037132

toclassi cationtechniquesusinglocalneighborhoodsinordertoincludespatialinformationintoaclassi er.

However,unsupervisedimagesegmentationisachallengingtask.Segmentationaimsatdividinganimageintohomo-geneousregions,butthemeasureofhomogeneityisimagedependent[12].Dependingonthismeasure,theprocesscanresultinundersegmentation(severalregionsaredetectedasone)oroversegmentation(oneregionisdetectedasseveralones)oftheimage.Inpreviousworks[13],[14],wepreferredoversegmentationtoundersegmentationinordernottomissobjectsintheclassi cationmap.Inthiswork,weaimtoreduceoversegmentationandthusfurtherimprovesegmentationandclassi cationresults.Thiscanbeachievedbyusingmarkersorregionseeds[12],[15].Inpreviousstudies,amarker(aninternalmarker)wasde nedasaconnectedcomponentbelong-ingtotheimageandassociatedwithanobjectofinterest[12],[15]–[17].Inourstudy,wede neamarkerasasetofimagepixels(notnecessarilyconnected;itcanbecomposedofseveralspatiallydisjointsubsetsofadjacentpixels)whichisassociatedwithoneobjectintheimagescene.

Theproblemofautomaticmarkerselectionhaspreviouslybeendiscussedintheliterature,mostlyforgray-scaleandcolorimages.Markersareoftende nedbysearching atzones(i.e.,connectedcomponentsofpixelsofconstantgray-levelvalue),zonesofhomogeneoustexture,orimageextrema[15].Gómezetal.[18]appliedhistogramanalysistoobtainasetofrepresentativepixelvalues,andthemarkersweregener-atedwithalltheimagepixelswithrepresentativegrayvalues.Jalbaetal.[16]usedconnectedoperators lteringonthegradientimageinordertoselectmarkersforagray-scalediatomimage.Noyeletal.[17],[19]performedclassi cationofthehyperspectralimage(usingdifferenttechniques,suchasClara[20]andlineardiscriminantanalysis)andthen lteredtheclassi cationmapsclassbyclass,usingmorphologicaloperators,inordertoselectlargespatialregionsasmarkers.Furthermore,theauthorsproposedtouserandomballs(con-nectedsetsofpixelsofrandomlyselectedsizes)extractedfromtheselargeregionsasmarkers.Inthediscussedstudies[16],[17],[19],theobjectivewastosegmentspeci cstructures(bloodcells,diatoms,glueocclusions,andcancerousgrowth).Inourstudy,theobjectiveistomark(selectamarkerfor)eachsigni cantspatialobjectintheimage.Here,bysigni cant,wemeananobjectofatleastone-pixelsizethatbelongstooneoftheclassesofinterest.Asremotesensingimagescontainsmallandcomplexstructures,automaticselectionofmarkersisanespeciallychallengingtask.

Inthispaper,anewschemeformarker-basedsegmenta-tionandclassi cationofhyperspectralimagesisproposed.Inparticular,weproposetoperformaprobabilisticpixelwiseclassi cation rstinordertochoosethemostreliableclassi edpixelsasmarkersofspatialregions[21].Furthermore,imagepixelsaregroupedintoaminimumspanningforest(MSF)[22],whereeachtreeisrootedonaclassi cation-derivedmarker.Thedecisiontoconnectthepixel,whichisnotyetintheforest,tooneofthetreesintheforestisbasedonitssimilaritytooneoftheadjacentpixelsalreadybelongingtotheforest.Byassigningaclassofthemarkertoallthepixelswithintheregiongrownfromtheconsideredmarker,aspectral-

spatialclassi cationmapisobtained.Furthermore,theclassi- cationmapisre nedusingtheresultsofapixelwiseclassi- cationandamajorityvotingwithinthes …… 此处隐藏：43651字，全部文档内容请下载后查看。喜欢就下载吧 ……