compositionTRODisusedtorepresentaclassofimagesbasedonvariancemaximizationandgreedysuccessiveresiduecalculationAtwodimensionalPCA2DPCAisproposedinYangetal2004thatconstructsanimagecovaria ID: 203373
Download Pdf The PPT/PDF document "UncorrelatedMultilinearPrincipalComponen..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
UncorrelatedMultilinearPrincipalComponentAnalysis composition(TROD)isusedtorepresentaclassofim-agesbasedonvariancemaximizationand(greedy)suc-cessiveresiduecalculation.Atwo-dimensionalPCA(2DPCA)isproposedin(Yangetal.,2004)thatcon-structsanimagecovariancematrixusingimagema-tricesasinputs.However,lineartransformationisappliedonlytotherightsideofimagematricessotheimagedataisprojectedinonemodeonly,result-inginpoordimensionalityreduction.Amoregeneralalgorithmnamedgeneralizedlowrankapproximationofmatrices(GLRAM)wasintroducedin(Ye,2005),whichappliestwolineartransformstoboththeleftandrightsidesofinputimagematricesandresultsinabetterdimensionalityreductionthan2DPCA.GLRAMisdevelopedfromtheperspectiveofapprox-imationwhilethegeneralizedPCA(GPCA)ispro-posedin(Yeetal.,2004)fromtheviewofvariationmaximization,asanextensionofPCA.Later,thecon-currentsubspacesanalysis(CSA)isformulatedin(Xuetal.,2005)foroptimalreconstructionofgeneralten-sorobjects,whichcanbeconsideredasageneraliza-tionofGLRAM,andthemultilinearPCA(MPCA)introducedin(Luetal.,2008a)targetsatvariationmaximizationforgeneraltensorobjectsintheexten-sionofPCAtothemultilinearcase,whichcanbecon-sideredasafurthergeneralizationofGPCA.However,noneoftheexistingmultilinearextensionsofPCAmentionedabovetakesanimportantpropertyofPCAintoaccount,i.e.,PCAderivesuncorrelatedfeatures,whichcontainminimumredundancyanden-sureindependenceamongfeatures.Instead,mostofthemproduceorthogonalbasesineachmode.Al-thoughuncorrelatedfeaturesimplyorthogonalprojec-tionbasesinPCA,thisisnotnecessarilythecaseforitsmultilinearextension.Withthismotivation,thispaperinvestigatesmultilinearextensionofPCAthatcanproduceuncorrelatedfeatures.WeproposeanoveluncorrelatedmultilinearPCA(UMPCA)forunsuper-visedtensorobjectdimensionalityreduction(featureextraction).UMPCAisbasedonthetensor-to-vectorprojection(TVP)(Luetal.,2008b)anditfollowstheclassicalPCAderivationofsuccessivevariancemaxi-mization(Jollie,2002).Thus,anumberofelemen-tarymultilinearprojections(EMPs)aresolvedtomax-imizethecapturedvariancewiththezero-correlationconstraint.Thesolutionisiterativeinnature,asmanyothermultilinearalgorithms(Xuetal.,2005;Yeetal.,2004;Shashua&Levin,2001).Therestofthispaperisorganizedasfollows.Section2reviewsbasicmultilinearnotationsandoperations,aswellastheconceptoftensor-to-vectorprojection.InSec.3,theproblemofUMPCAisformulatedandthesolutionisderivedasasequentialiterativeprocess.Table1.Notations NotationsDescriptions Xm,m=1;:::;Mthemthinputtensorsampleu(n),n=1;:::;Nthen-modeprojectionvectorfu(n)Tp;n=1;:::;NgthepthEMP,wherepistheindexoftheEMPymtheprojectionofXmontheTVPfu(n)Tp;n=1;:::;NgPp=1ym(p)=ymp=gp(m)theprojectionofXmonthepthEMPfu(n)Tp;n=1;:::;Nggpthepthcoordinatevector Next,Sec.4evaluatestheeectivenessofUMPCAinthepopularfacerecognitiontaskthroughcomparisonwithPCA,MPCAandTROD.Finally,theconclusionsaredrawninSec.5.2.MultilinearFundamentalsThissectionintroducesthemultilinearnotations,op-erationsandprojectionsneededinthepresentationofUMPCA,andforfurtherpursuingofmultilinearalge-bra,(Lathauweretal.,2000)isagoodreference.TheimportantnotationsusedinthispaperarelistedinTable1forhandyreference.2.1.NotationsandbasicmultilinearoperationsDuetothemultilinearnatureoftensorobjects,newnotationshavebeenintroducedintheliteratureformathematicalanalysis.Followingthenotationsin(Lathauweretal.,2000),wedenotevectorsbylow-ercaseboldfaceletters,e.g.,x;matricesbyuppercaseboldfaceletters,e.g.,U;andtensorsbycalligraphicletters,e.g.,A.Theirelementsaredenotedwithin-dicesinparentheses.Indicesaredenotedbylowercaselettersandspantherangefrom1totheuppercaseletteroftheindex,e.g.,n=1;2;:::;N.AnNth-ordertensorA2RI1I2:::INisaddressedbyNindicesin,n=1;:::;N,andeachinaddressesthen-modeofA.Then-modeproductofatensorAbyamatrixU2RJnIn,denotedbyAnU,isatensorwithentries:(AnU)(i1;:::;in1;jn;in+1;:::;iN)=XinA(i1;i2;:::;iN)U(jn;in):(1)ThescalarproductoftwotensorsA;B2RI1I2:::INisdenedas:A;B]TJ/;༔ ; .96;& T; 10;.516; 0 T; [0;=Xi1:::XiNA(i1;:::;iN)B(i1;:::;iN):(2)Arank-onetensorAequalstotheouterproductofN UncorrelatedMultilinearPrincipalComponentAnalysis Inotherwords,theUMPCAobjectiveistodetermineasetofPEMPsfu(n)Tp;n=1;:::;NgPp=1thatmaxi-mizethevariancewhileproducingfeatureswithzero-correlation.Thus,theobjectivefunctionforthepthEMPisfu(n)Tp;n=1;:::;Ng=argmaxMXm=1(ymp yp)2;subjecttou(n)Tpu(n)p=1andgTpgq kgpkkgqk=pq;p;q=1;:::;P;(6)wherepqistheKroneckerdelta(denedas1forp=qandas0otherwise).3.2.TheUMPCAalgorithmTosolvetheUMPCAproblem(6),wefollowthesuc-cessivevariancemaximizationapproachinthederiva-tionofPCAin(Jollie,2002).ThePEMPsfu(n)Tp;n=1;:::;NgPp=1aredeterminedonebyoneinPsteps,withthepthstepobtainingthepthEMP:Step1:DeterminetherstEMPfu(n)T1;n=1;:::;NgbymaximizingSyT1withoutanycon-straint.Step2:DeterminethesecondEMPfu(n)T2;n=1;:::;NgbymaximizingSyT2subjecttothecon-straintthatgT2g1=0.Stepp(p=3;:::;P):DeterminethepthEMPfu(n)Tp;n=1;:::;NgbymaximizingSyTpsubjecttotheconstraintthatgTpgq=0forq=1;:::;p1.InordertosolveforthepthEMPfu(n)Tp;n=1;:::;Ng,weneedtodetermineNsetsofparameterscorrespond-ingtoNprojectionvectors,u(1)p;u(2)p;:::u(N)p,oneineachmode.Unfortunately,simultaneousdetermina-tionoftheseNsetsofparametersinallmodesisacomplicatednon-linearproblemwithoutanexistingoptimalsolution,exceptwhenN=1,whichistheclassicalPCAwhereonlyoneprojectionvectoristobesolved.Therefore,wefollowtheapproachinthealternatingleastsquare(ALS)algorithm(Harshman,1970)tosolvethismultilinearproblem.ForeachEMPtobedetermined,theparametersoftheprojectionvec-toru(n)pforeachmodenareestimatedonemodebyonemodeseparately,conditionedonfu(n)p;n6=ng,theparametervaluesoftheprojectionvectorsintheothermodes.Tosolveforu(n)pinthen-mode,assumingthatfu(n)p;n6=ngisgiven,thetensorsamplesarepro-jectedinthese(N1)modesfn6=ngrsttoobtainthevectors~y(n)mp=Xm1u(1)Tp:::n1u(n1)Tpn+1u(n+1)Tp:::Nu(N)Tp;(7)where~y(n)mp2RIn.Thisconditionalsubproblemthenbecomestodetermineu(n)pthatprojectsthevectorsamplesf~y(n)mp;m=1;:::;Mgontoalinesothatthevarianceismaximized,subjecttothezero-correlationconstraint,whichisaPCAproblemwiththeinputsamplesf~y(n)mp;m=1;:::;Mg.Thecorrespondingto-talscattermatrix~S(n)Tpisthendenedas~S(n)Tp=MXm=1(~y(n)mp~y(n)p)(~y(n)mp~y(n)p)T;(8)where~y(n)p=1 MPm~y(n)mp.With(8),wearereadytosolveforthePEMPs.Forp=1,theu(n)1thatmaximizesthetotalscatteru(n)T1~S(n)T1u(n)1intheprojectedspaceisobtainedastheuniteigenvectorof~S(n)T1associatedwiththelargesteigenvalue.Next,weshowhowtodeterminethepth(p1)EMPgiventherst(p1)EMPs.Giventherst(p1)EMPs,thepthEMPaimstomaximizethetotalscatterSyTp,subjecttotheconstraintthatfeaturesprojectedbythepthEMPareuncorrelatedwiththoseprojectedbytherst(p1)EMPs.Let~Y(n)p2RInMbeamatrixwith~y(n)mpasitsmthcolumn,i.e.,~Y(n)p=h~y(n)1p;~y(n)2p;:::;~y(n)Mpi,thenthepthcoordinatevectorisgp=~Y(n)Tpu(n)p.Theconstraintthatgpisun-correlatedwithfgq;q=1;:::;p1gcanbewrittenasgTpgq=u(n)Tp~Y(n)pgq=0;q=1;:::;p1:(9)Thus,u(n)p(p1)canbedeterminedbysolvingthefollowingconstrainedoptimizationproblem:u(n)p=argmaxu(n)Tp~S(n)Tpu(n)p;(10)subjecttou(n)Tpu(n)p=1andu(n)Tp~Y(n)pgq=0;q=1;:::;p1;Thesolutionisgivenbythefollowingtheorem:Theorem1.Thesolutiontotheproblem(10)isthe(unit-length)eigenvectorcorrespondingtothelargesteigenvalueofthefollowingeigenvalueproblem: (n)p~S(n)Tpu=u;(11) UncorrelatedMultilinearPrincipalComponentAnalysis 3.3.Initialization,projectionorderandterminationAsaniterativealgorithm,theUMPCAmaybeaf-fectedbytheinitializationmethod,theprojectionor-derandtheterminationconditions.Duetothespaceconstraint,theseissues,aswellastheconvergenceandcomputationalissues,arenotstudiedhere.Instead,weadoptsimpleimplementationstrategiesforthem.First,weusetheuniforminitializationforUMPCA,wherealln-modeprojectionvectorsareinitializedtohaveunitlengthandthesamevaluealongtheIndi-mensionsinn-mode,whichisequivalenttotheallonesvector1withpropernormalization.Second,asshowninAlgorithm1,theprojectionorder,whichisthemodeorderingincomputingtheprojectionvectors,isfrom1-modetoN-mode,asinothermultilinearalgorithms(Ye,2005;Xuetal.,2005;Luetal.,2008a).Third,theiterationisterminatedbysettingK,themaximumnumberofiterations.4.ExperimentalEvaluationTheproposedUMPCAcanpotentiallybenetvariousapplicationsinvolvingtensorialdata,asmentionedinSec.1.Sincefacerecognitionhaspracticalimpor-tanceinsecurity-relatedapplicationssuchasbiomet-ricauthenticationandsurveillance,ithasbeenusedwidelyforevaluationofunsupervisedlearningalgo-rithms(Shashua&Levin,2001;Yangetal.,2004;Xuetal.,2005;Ye,2005).Therefore,inthissection,wefocusonevaluatingtheeectivenessofUMPCAonthispopularclassicationtaskthroughperformancecomparisonwithexistingunsuperviseddimensionalityreductionalgorithms.4.1.TheFERETdatabaseTheFacialRecognitionTechnology(FERET)database(Phillipsetal.,2000)iswidelyusedfortestingfacerecognitionperformance,with14,126imagesfrom1,199subjectscoveringawiderangeofvariationsinviewpoint,illumination,facialex-pression,racesandages.Asubsetofthisdatabaseisselectedinourexperimentalevaluationanditconsistsofthosesubjectswitheachsubjecthavingatleasteightimageswithatmost15degreesofposevariation,resultingin721faceimagesfrom70subjects.Sinceourfocushereisontherecognitionoffacesratherthantheirdetection,allfaceimagesaremanuallycropped,aligned(withmanuallyannotatedcoordinateinformationofeyes)andnormalizedto8080pixels,with256graylevelsperpixel.Figure1showssomesamplefaceimagesfromtwosubjectsinthisFERETsubset. Figure1.ExamplesoffaceimagesfromtwosubjectsintheFERETsubsetusedinourexperimentalevaluation.4.2.FacerecognitionperformancecomparisonIntheevaluation,wecomparetheperformanceoftheUMPCAagainstthreePCA-basedunsupervisedlearningalgorithms:thePCA(eigenface)algorithm(Turk&Pentland,1991),theMPCAalgorithm(Luetal.,2008a)2andtheTRODalgorithm(Shashua&Levin,2001).ThenumberofiterationsinTRODandUMPCAissettoten,withthesame(uniform)initial-izationused.ForMPCA,weobtainthefullprojectionandselectthemostdescriptivePfeaturesforrecogni-tion.Thefeaturesobtainedbythesefouralgorithmsarearrangedindescendingvariationcaptured(mea-suredbyrespectivetotalscatter).Forclassicationofextractedfeatures,weusethenearestneighborclassi-er(NNC)withEuclideandistancemeasure.Gray-levelfaceimagesarenaturallysecond-orderten-sors(matrices),i.e.,N=2.Therefore,theyareinputdirectlyas8080tensorstothemultilin-earalgorithms(MPCA,TROD,UMPCA),whileforPCA,theyarevectorizedto64001vectorsasin-put.Foreachsubjectinafacerecognitionexperiment,L(=1;2;3;4;5;6;7)samplesarerandomlyselectedforunsupervisedtrainingandtherestareusedfortesting.Wereporttheresultsaveragedovertensuchrandomsplits(repetitions).Figures2and3showthedetailedresults3forL=1andL=7,respectively.L=1isanextremesmallsamplesizescenariowhereonlyonesampleperclassisavailablefortraining,theso-calledonetrainingsample(OTS)caseimportantinpractice(Wangetal.,2006),andL=7isthemaximumnumberoftrainingsampleswecanuseinourexperiments.Figures2(a)and3(a)plotthecorrectrecognitionratesagainstP,thedi-mensionalityofthesubspaceforP=1;:::;10,andFigs2(b)and3(b)plotthoseforP=15;:::;80.Fromthegures,UMPCAoutperformstheotherthreemethodsinbothcasesandacrossalldimensionality,indicatingthattheuncorrelatedfeaturesextracteddirectlyfromthetensorialfacedataaremoreeectiveinclassi- 2NotethatMPCAwithN=2isequivalenttoGPCA.3NotethatforPCAandUMPCA,thereareatmost69featureswhenL=1(only70facesfortraining). UncorrelatedMultilinearPrincipalComponentAnalysis 5.ConclusionsThispaperproposesanoveluncorrelatedmultilinearPCAalgorithm,whereuncorrelatedfeaturesareex-tracteddirectlyfromtensorialrepresentationthroughatensor-to-vectorprojection.Thealgorithmsucces-sivelymaximizesvariancecapturedbyeachelemen-taryprojectionwhileenforcingthezero-correlationconstraint.Thesolutionemploysthealternatingpro-jectionmethodandisiterative.ExperimentsonfacerecognitiondemonstratethatcomparedwithotherunsupervisedlearningalgorithmsincludingthePCA,MPCAandTROD,theUMPCAachievesthebestre-sultsanditisparticularlyeectiveinlow-dimensionalspaces.Thus,facerecognitionthroughunsupervisedlearningbenetsfromtheproposedUMPCAandinfutureresearch,itisworthwhiletoinvestigatewhetherUMPCAcancontributeinotherunsupervisedlearningtasks,suchasclustering.AcknowledgmentsTheauthorswouldliketothanktheanonymousre-viewersfortheirinsightfulcomments.ThisworkispartiallysupportedbytheOntarioCentresofExcel-lencethroughtheCommunicationsandInformationTechnologyOntarioPartnershipProgramandtheBellUniversityLabs-attheUniversityofToronto.ReferencesCarroll,J.D.,&Chang,J.J.(1970).Analysisofindi-vidualdierencesinmultidimensionalscalingviaann-waygeneralizationof\eckart-young"decomposi-tion.Psychometrika,35,283{319.Faloutsos,C.,Kolda,T.G.,&Sun,J.(2007).Min-inglargetime-evolvingdatausingmatrixandtensortools.Int.Conf.onDataMining2007Tutorial.Harshman,R.A.(1970).Foundationsoftheparafacprocedure:Modelsandconditionsforan\explana-tory"multi-modalfactoranalysis.UCLAWorkingPapersinPhonetics,16,1{84.Jollie,I.T.(2002).Principalcomponentanalysis,secondedition.SpringerSeriresinStatistics.Koren,Y.,&Carmel,L.(2004).Robustlineardi-mensionalityreduction.IEEETrans.Vis.Comput.Graphics,10,459{470.Lathauwer,L.D.,Moor,B.D.,&Vandewalle,J.(2000).Onthebestrank-1andrank-(R1;R2;:::;RN)approximationofhigher-orderten-sors.SIAMJournalofMatrixAnalysisandAppli-cations,21,1324{1342.Law,M.H.C.,&Jain,A.K.(2006).Incrementalnon-lineardimensionalityreductionbymanifoldlearn-ing.IEEETrans.PatternAnal.Mach.Intell.,28,377{391.Lu,H.,Plataniotis,K.N.,&Venetsanopoulos,A.N.(2008a).MPCA:Multilinearprincipalcomponentanalysisoftensorobjects.IEEETrans.NeuralNetw.,19,18{39.Lu,H.,Plataniotis,K.N.,&Venetsanopoulos,A.N.(2008b).Uncorrelatedmultilineardiscriminantanalysiswithregularizationandaggregationforten-sorobjectrecognition.IEEETrans.NeuralNetw.acceptedpendingminorrevision.Phillips,P.J.,Moon,H.,Rizvi,S.A.,&Rauss,P.(2000).TheFERETevaluationmethodforfacerecognitionalgorithms.IEEETrans.PatternAnal.Mach.Intell.,22,1090{1104.Shakhnarovich,G.,&Moghaddam,B.(2004).Facerecognitioninsubspaces.HandbookofFaceRecog-nition(pp.141{168).Springer-Verlag.Shashua,A.,&Levin,A.(2001).Linearimagecod-ingforregressionandclassicationusingthetensor-rankprinciple.Proc.IEEEConf.onComputerVi-sionandPatternRecognition(pp.42{49).Turk,M.,&Pentland,A.(1991).Eigenfacesforrecog-nition.JournalofCognitiveNeurosicence,3,71{86.Wang,J.,Plataniotis,K.N.,Lu,J.,&Venetsanopou-los,A.N.(2006).Onsolvingthefacerecognitionproblemwithonetrainingsamplepersubject.Pat-ternRecognition,39,1746{1762.Xu,D.,Yan,S.,Zhang,L.,Zhang,H.-J.,Liu,Z.,&Shum;,H.-Y.(2005).Concurrentsubspacesanaly-sis.Proc.IEEEComputerSocietyConf.onCom-puterVisionandPatternRecognition(pp.203{208).Yang,J.,Zhang,D.,Frangi,A.F.,&Yang,J.(2004).Two-dimensionalPCA:anewapproachtoappearance-basedfacerepresentationandrecogni-tion.IEEETrans.PatternAnal.Mach.Intell.,26,131{137.Ye,J.(2005).Generalizedlowrankapproximationsofmatrices.MachineLearning,61,167{191.Ye,J.,Janardan,R.,&Li,Q.(2004).GPCA:Ane-cientdimensionreductionschemeforimagecom-pressionandretrieval.The10thACMSIGKDDInt.Conf.onKnowledgeDiscoveryandDataMin-ing(pp.354{363).