资源描述:
《Kernel Discriminant Learning with Application to face recognition》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、KernelDiscriminantLearningwithApplicationtoFaceRecognitionJuweiLu1,K.N.Plataniotis2,andA.N.Venetsanopoulos3BellCanadaMultimediaLaboratoryTheEdwardS.RogersSr.DepartmentofElectricalandComputerEngineeringUniversityofToronto,Toronto,M5S3G4,Ontario,Canada123Emails:{juwei,kostas,anv}@dsp.
2、toronto.eduAbstract.Whenappliedtohigh-dimensionalpatternclassificationtaskssuchasfacerecognition,traditionalkerneldiscriminantanalysismethodsoftensufferfromtwoproblems:1)smalltrainingsamplesizecomparedtothedimensionalityofthesample(ormappedkernelfeature)space,and2)highcomputationalcom
3、plexity.Inthischapter,weintroduceanewkerneldiscriminantlearningmethod,whichattemptstodealwiththetwoproblemsbyusingregularizationandsubspacede-compositiontechniques.Theproposedmethodistestedbyextensiveexperimentsperformedonrealfacedatabases.Theobtainedresultsindicatethatthemethodoutp
4、erforms,intermsofclassificationaccuracy,existingkernelmethods,suchaskernelPrincipalComponentAnalysisandkernelLinearDiscriminantAnalysis,atasignificantlyreducedcomputationalcost.Keywords:StatisticalDiscriminantAnalysis,KernelMachines,SmallSampleSize,NonlinearFeatureExtraction,FaceRecog
5、nition1IntroductionStatisticallearningtheorytellsusessentiallythatthedifficultyofanesti-mationproblemincreasesdrasticallywiththedimensionalityJofthesamplespace,sinceinprinciple,asafunctionofJ,oneneedsexponentiallymanypat-ternstosamplethespaceproperly[18,32].Unfortunately,inmanypractic
6、altaskssuchasfacerecognition,thenumberofavailabletrainingsamplespersubjectisusuallymuchsmallerthanthedimensionalityofthesamplespace.Forinstance,acanonicalexampleusedforfacerecognitionisa112×92image,whichexistsina10304-dimensionalrealspace.Nevertheless,thenumberofexamplesperclassavai
7、lableforlearningisnotmorethanteninmostcases.Thisresultsintheso-calledsmallsamplesize(SSS)problem,whichisknowntohavesignificantinfluencesontheperformanceofastatisticalpatternrecog-nitionsystem(seee.g.[3,5,9,12,13,16,21,33,34]).WhenitcomestostatisticaldiscriminantlearningtaskssuchasLine
8、arDis-criminantAnal