简介:WhenusingAdaBoosttoselectdiscriminantfeaturesfromsomefeaturespace(e.g.Gaborfeaturespace)forfacerecognition,cascadestructureisusuallyadoptedtoleveragetheasymmetryinthedistributionofpositiveandnegativesamples.EachnodeinthecascadestructureisaclassifiertrainedbyAdaBoostwithanasymmetriclearninggoalofhighrecognitionratebutonlymoderatelowfalsepositiverate.OnelimitationofAdaBoostarisesinthecontextofskewedexampledistributionandcascadeclassifiers:AdaBoostminimizestheclassificationerror,whichisnotguaranteedtoachievetheasymmetricnodelearninggoal.Inthispaper,weproposetousetheasymmetricAdaBoost(Asym-Boost)asamechanismtoaddresstheasymmetricnodelearninggoal.Moreover,thetwopartsoftheselectingfeaturesandformingensembleclassifiersaredecoupled,bothofwhichoccursimultaneouslyinAsymBoostandAdaBoost.FisherLinearDiscriminantAnalysis(FLDA)isusedontheselectedfea-turestolearnalineardiscriminantfunctionthatmaximizestheseparabilityofdataamongthedifferentclasses,whichwethinkcanimprovetherecognitionperformance.Theproposedalgorithmisdem-onstratedwithfacerecognitionusingaGaborbasedrepresentationontheFERETdatabase.Ex-perimentalresultsshowthattheproposedalgorithmyieldsbetterrecognitionperformancethanAdaBoostitself.
简介:Anoveltextindependentspeakeridentificationsystemisproposed.Intheproposedsystem,the12-orderperceptuallinearpredictivecepstrumandtheirdeltacoefficientsinthespanoffiveframesareextractedfromthesegmentedspeechbasedonthemethodofpitchsynchronousanalysis.TheFisherratiosoftheoriginalcoefficientsthenbecalculated,andthecoefficientswhoseFisherratiosarebiggerareselectedtoformthe13-dimensionalfeaturevectorsofspeaker.TheGaussianmixturemodelisusedtomodelthespeakers.TheexperimentalresultsshowthattheidentificationaccuracyoftheproposedsystemisobviouslybetterthanthatofthesystemsbasedonotherconventionalcoefficientslikethelinearpredictivecepstralcoefficientsandtheMel-frequencycepstralcoefficients.