Zhang_Sparse_Representation_Classification_2015_CVPR_paper

《Zhang_Sparse_Representation_Classification_2015_CVPR_paper》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、SparseRepresentationClassificationwithManifoldConstraintsTransferBaochangZhang;yAlessandroPerinaVittorioMurinoAlessioDelBueIstitutoItalianodiTecnologia(IIT)PatternAnalysisandComputerVision(PAVIS)ViaMorego30,16136Genova,ItalyySchoolofAutomationScienceandElectricalEnginee

2、ringBeihangUniversity,Beijing,ChinaAbstractThefactthatimagedatasampleslieonamanifoldhasbeensuccessfullyexploitedinmanylearningandinferenceproblems.Inthispaperweleveragethespecificstructureofdatainordertoimproverecognitionaccuraciesingen-eralrecognitiontasks.Inparticularwepro

3、poseanovelframeworkthatallowstoembedmanifoldpriorsintosparserepresentation-basedclassification(SRC)approaches.WealsoshowthatmanifoldconstraintscanbetransferredfromFigure1.AschemeoftheMCTapproach.thedatatotheoptimizedvariablesifthesearelinearlycor-related.Usingthisnewinsight,

4、wedefineanefficiental-ternatingdirectionmethodofmultipliers(ADMM)thatcaninstantiatesrecognitionasasparseoptimizationproblemconsistentlyintegratethemanifoldconstraintsduringthewheretheaimistoestimateanassignmentmatrixZthatas-optimizationprocess.Thisisbasedonthepropertythatsoci

5、ateseachtestingsampleHrtothecorrespondingtrain-wecanrecasttheproblemastheprojectionoverthemani-ingdataHdsuchthat:foldviaalinearembeddingmethodbasedontheGeodesicdistance.TheproposedapproachissuccessfullyappliedonHr=HdZ:face,digit,actionandobjectsrecognitionshowingaconsis-Thi

6、sinferenceisingeneraldonethroughtheuseofadictio-tentlyincreaseonperformancewhencomparedtothestatenarycomputeddirectlyfromthetrainingdata.Ifsuchdic-oftheart.tionaryhascertainproperties,namelybeinguncorrelated,thetheoryofSRChasinterestinglinkstothecompres-1.Introductionsivese

7、nsingframework.Thismakestheproblemsolvablethroughal1relaxationthatcanbedealtwithefficientopti-Aconstrainedlearningmodelallowsonetoincorporatemizationtechniques.domain-specificknowledgeasconstraintstobalancetheThispaperprovidesnewinsightstotheproblem,inpar-learnedmodelgiventhe

8、implicitstructureofthedata[7].ticulararelevantintuitionthatwasneglectedinpreviousF