基于流形學習與子空間的降維方法研究與應用.pdf

1、東北師范大學碩士學位論文基于流形學習與子空間的降維方法研究與應用姓名：劉迪申請學位級別：碩士專業(yè)：計算機應用技術指導教師：呂英華20090501AbstractTheimprovingabilitiesofdatacollectionandstoragecapabilitiesduringthepastdecadeshaveledtoaninf6nnationoverloadinmostscientificdomainsTraditio

2、nalalgorithmsusedinmachinelearningandpatternrecognitionapplicationsateoftensusceptibletothewell’knownproblemofthecurseofdimensionality，whichreferstothedegradationintheperformanceofagivenlearningalgorithmasthenumberoffeat

3、uresincreasesTodealwiththisissue，dimensionalityreductiontechniquesateoftenappliedasadatapre。processmgsteporaspartofthedataanalysistosimplifythedatamodelDimensionalityreductionisthetransformationofhighdimensionaldataintoa

4、meaningfulrepresentationofreduceddimensionalityItisimportantinmanydomains，sinceitappliedmclassificationvisualization，andcompression“ofhigh—dimensionaldataWhenthesuitablelowdimensionalsubspaceisidentified，whichrepresentfo

5、rtheoriginalhigh。dimensionaldatasetthroughthedimensionalityreductiontransformation，taskssuchasclassificationorclusteringCanoftenyieldmoreaccurateandreadilyinterpretableresults，whilecomputationalcostsmayalsobesignificantl

6、yreducedThedissertationpresentsareviewandcomparativestudyoftechniquesfordimensionalityreductionatpresentWealsopointoutthedisadvantageofthesemethodsandthenproposeanoveldimensionalityreductionmethodbasedonmanifoldlearninga

7、ndsubspace，DRMSItCanpreserveintraclassneighboringgeometrystructureandextractbetween‘classrelevantstructuresforclassificationeffectivelyTheproposedmethodisappliedtomultimodebiometricsrecognitionsystemwithfaceandpalmpfinti

8、magesThemethodisexaminedusingtheORLandFERETfacedatabasesandPolvUpalmprintdatabaseExperimentalresultsshowthatDRMSconsistentlyoutperformsotherlineardimensionalityreductionmethodswhenthetrainingsamplesizeofperclassISsmallTh