基于统计学习理论的支持向量机分类方法研究-信号与信息处理专业论文.docx

中国科学技术大学博士学位论文AbstractAbstractAsalimited—samplelearningtheory,StatisticalLearningTheory(SLT)hasmanyadvantagesinpatternrecognition，suchasitssuperiorityinsmall—sample，nonlinearandhigh—dimensionproblems．Itreinforcesthetheoryofpatternrecognitionandmachinelearning．SupportVectorMachine(SVM)isoneofthelearningmethodsthatdevelopedfromSLT．Whendealwiththesmall-samplelearning，itcomestheoptimalsolutionofthelimitedinformationandsolvesmanyproblemssuchasmodelselection，overfitting，nonlinear,dimensiondisasterinhighdegree．Now)manyresearchesfocusonSLTandSVMinpatternrecognitionfield．ThispaperresearchesthealgorithmsandapplicationsofSVMclassifierbasedonSET,analysestheproblemsuchaskernelproblem，computespeedproblem，designofmulti-classSVMandfacedetectionwithSVM，anddevelopssomeeffectiveandefficientsolutions．Themaincontentsandcontributionsofthisdissertationliein6aspects．1．Anewsupportvectormachinekernelbasedonspatialdistance——————MDGK(Moderate—DecreasingGaussKernel)isproposed．It’SprovedthattheMDGKkernelmakeslessSVs，widermarginandhighercorrectnessthanGaussandExponentialRBFkernel．Therefore，thetwoclassesCanbet’separatedwithwidermarginandbettergeneralization．2．AnewmethodofusingthemixedkernelsintheSVMclassifiertogetbetterperformanceofclassifyisdeveloped．StudiedcharacteristicsofdifferentkemeI譬——一globalkernelandlocalkernel．AndanewmethodofusingthemixedkernelsintheSVMclassifierisproposedtogetbetterperformanceofclassify,especiallythegeneralizationability．3．AnewGeometricfastalgorithmofsupportvectormachinesandtheconceptof‘hullveetor”areproposed．Usingthegeometricinformationofthetrainingsamples，thealgorithmfirstextractsthesetofhullvectors，whicharemostlikelytobethesupportvector．ThenthesetofhullvectorsaretrmnedasthenewtrainingsamplestoIII!里型兰垫查盔兰堡主兰垡堡苎——一——兰堕!l旦getthesupportvectors．HVSVMreducesthetimeconsumedbytheQPproblemintheSVMtraininginlargedegree，andhighlyspeedsthewholetrainingprocessofSVM．AndthismethoddoesnotdeclinetheperformanceofSVM．4．AnewSVMIterativeAlgorithmBasedonClusteringisproposed．Firstly,itchoos