您好,欢迎访问三七文档
MultipleClassifierSystemsforMultisensorDataFusionRobiPolikar*,DeviParikh§andShreekanthMandayamElectricalandComputerEngineering,RowanUniversity,Glassboro,NJ08028E-Mail:polikar@rowan.edu,dparikh@andrew.cmu.edu,shreek@rowan.edu*Contributingauthor:R.Polikar,136RowanHall,201MullicaHillRoad,Glassboro,NJ08028~polikar/RESEARCH§D.ParikhisnowwiththeElectricalandComputerEngineering,CarnegieMellonUniversity,Pittsburgh,PA15123Abstract–WehavepreviouslyintroducedLearn++,anen-sembleofclassifiersbasedalgorithmcapableofincre-mentallearningfromadditionaldata,andpointedtoitsfea-sibilityindatafusionapplications.Inthiscontribution,weprovideadditionaldetails,updatedresultsandinsightonhowsuchasystemcanbeusedinintegratingcomplemen-taryknowledgeprovidedbydifferentdatasourcesobtainedfromdifferentsensors.Essentially,thealgorithmgeneratesanensembleofclassifiersusingdatafromeachsource,andcombinestheseclassifiersusingaweightedvotingproce-dure.Theweightsaredeterminedbasedontheindividualclassifier’strainingperformanceaswellastheobservedorpredictedreliabilityofeachdatasource.Keywords-Fusion,combiningclassifiers,ensemblesys-tems,incrementallearning,Learn++I.INTRODUCTIONInmanyapplicationsofpatternrecognitionandauto-matedidentification,itisnotuncommonfordataobtainedfromdifferentsensorsmonitoringaphysicalphenomenontoprovidecomplimentaryinformation.Asuitablecombinationofsuchinformationisusuallyreferredtoasdataorinfor-mationfusion,andcanleadtoimprovedaccuracyandcon-fidenceoftheclassificationdecisioncomparedtoadecisionbasedonanyoftheindividualdatasourcesalone.WehavepreviouslyintroducedLearn++,anensembleofclassifiersbasedapproach,asaneffectiveautomatedclassi-ficationalgorithmthatiscapableoflearningincrementally.Thealgorithmiscapableofacquiringnovelinformationfromadditionaldatathatlaterbecomeavailableaftertheclassificationsystemhasalreadybeendesigned.Toachieveincrementallearning,Learn++generatesanensembleofclassifiers(experts),whereeachclassifieristrainedonthecurrentlyavailabledatabase.Recognizingtheconceptualsimilaritybetweendatafusionandincrementallearning,wediscussasimilarapproachfordatafusion:employanen-sembleofexperts,eachtrainedondataprovidedbyoneofthesources,andthenstrategicallycombinetheiroutputs.Wehaveobservedthattheperformanceofsuchasystemindecisionmakingapplicationsissignificantlyandconsis-tentlybetterthanthatofadecisionbasedonasingledatasourceacrossseveralbenchmarkandrealworlddatabases.Theapplicationsforsuchasystemarenumerous,wheredataavailablefrommultiplesources(ormultiplesensors)generatedbythesameapplicationmaycontaincomplemen-taryinformation.Forinstance,innon-destructiveevaluationofpipelines,defectinformationmaybeobtainedfromeddycurrent,magneticfluxleakageimages,ultrasonicscans,thermalimaging;ordifferentpiecesofdiagnosticinforma-tionmaybeobtainedfromseveraldifferentmedicaltests,suchasbloodanalysis,electrocardiographyorelectroe-ncephalography,medicalimagingdevices,suchasultra-sonic,magneticresonanceorpositronemissionscans,etc.Intuitively,ifsuchinformationfrommultiplesourcescanbeappropriatelycombined,theperformanceofaclassificationsystem(indetectingwhetherthereisadefect,orwhetheradiagnosticdecisioncanbemade)canbeimproved.Conse-quently,bothincrementallearninganddatafusioninvolvelearningfromdifferentsetsofdata.Inincrementallearningsupplementaryinformationmustbeextractedfromnewdatasets,whichmayincludeinstancesfromnewclasses.Indatafusion,complementaryinformationmustbeextractedfromnewdatasets,whichmayrepresentthedatausingdif-ferentfeatures.Traditionalmethodsaregenerallybasedonprobabilitytheory(Bayestheorem,Kalmanfiltering),ordecisiontheorysuchastheDempster-Schafer(DS)anditsvariations,whichwereprimarilydevelopedformilitaryapplications,suchasnotablytargetdetectionandtracking[1-3].Ensembleofclassifiersbasedapproachesseektoprovideafreshandamoregeneralsolutionforabroaderspectrumofapplica-tions.Itshouldalsobenotedthatinseveralapplications,suchasthenondestructivetestingandmedicaldiagnosticsmentionedabove,thedataobtainedfromdifferentsourcesmayhavebeengeneratedbydifferentphysicalmodalities,andthereforethefeaturesobtainedmaybeheterogeneous.WhileusingprobabilityordecisiontheorybasedapproachesSAS2006–IEEESensorsApplicationsSymposiumHouston,TexasUSA,7-9February20060-7803-9581-6/06/$20.00©2006IEEE180becomemorecomplicatedinsuchcases,heterogeneousfea-turescaneasilybeaccommodatedbyanensemblebasedsystem,asdiscussedbelow.Anensemblesystemcombinestheoutputsofseveraldi-verseclassifiersorexperts.Thediversityintheclassifiersallowsdifferentdecisionboundariestobegeneratedbyus-ingslightlydifferenttrainingparameters,suchasdifferenttrainingdatasets.Theintuitionisthateachexpertwillmakeadifferenterror,andstrategicallycombiningtheseclassifi-erscanreducetotalerror[4-6].Ensemblesystemshaveat-tractedagreatdealofattentionoverthelastdecadeduetotheirreportedsuperiorityoversingleclassifiersystemsonavarietyofapplications[7-10].Recognizingthepotentialofthisapproachforincre-mentallearningapplications,wehaverecentlydevelopedLearn++,andshownthatLearn++isindeedca
本文标题:Multiple Classifier Systems for Multisensor Data F
链接地址:https://www.777doc.com/doc-3283788 .html