基于Adaboost算法的人脸检测pdf

:2004-06-1422920059:1006-9348(2005)09-0167-03Adaboost,(,,200030):AdaboostAdaboost,,,Adaboost,,Adaboost,,PCA,,:;;;:TP391.41:AFaceDetectionBasedonAdaboostAlgorithmZHENGFeng,YANGXin(InstituteofImageProcessing&PatternRecognition,ShanghaiJiaotongUniversity,Shanghai200030,China)ABSTRACT:ThispaperpresentsanewalgorithmbasedonimprovedAdaboostforfacedetection.Adaboostisalearningalgorithmforconstructingaccurateclassifiers.Itcanobtainastronglearningalgorithmbycombiningaseriesofweaklearningalgorithmsthroughsomerules,butittendstooverfitinthepresenceofhighlynoisesamples.Inthispaper,weimprovetheweigh-updaterules,andusePCAandourimprovedAdaboosttodetectfaceimages.Experi2mentresultsshowthatthisapproachhasgoodperformance,andthemodifiedweigh-updaterulescaneffectivelya2voidoverfitting.KEYWORDS:Facedetection;Algorithm;PCA;Overfitting1,,[1,2],PaulViolaMichaelJones2001Adaboost[3],,,Adaboostboosting,,,;,,,Adaboost,,,,Adaboost,(overfitting),Adaboost,,PCA(),Adaboost,,,2Adaboost,KearnsValiant[4,5],:PAC(ProbablyApproxi2matelyCorrect),,,;,?761©1994-2007ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.[6]Adaboost(AdaptiveBoost),,Freund,Adaboost:{(x1,y1),(x2,y2),(xn,yn)},,xi,yi,,,yi{-1,+1},D=1n,T,,,(h1h2ht),,H:1):{(x1,y1),(x2,y2),,(xn,yn)},xi,xiX,X;yi,yiY={-1,+1}2):D1(xi)=1n,i=1n3)T,Fort=1,2,......,T:Dt,ht:X{-1,+1}:t=ni=1Dt(xi)[ht(xi)yi]t=12ln1-tt:Dt+1(xi)=Dt(xi)Zte-tht(xi)=yi,:yiht(xi)=+1etht(xi)yi,:yiht(xi)=-1=Dt(xi)exp(-tyiht(xi))Zt,Ztni=1Dt+1(xi)=1T,:H(x)=signnt=1tht(x)3Adaboost3.1[7]ErEl,:,,1,,,,,,,,13.2PCATurkPentland[8]KL,,NN,N2:x=(b11b21bN1b12b22bN2b1Nb2NbNN),,:=E{(x-)(x-)T}=1MM-1i=0{(xi-)(xi-)T}=1MXXT,xii,,M,X=[(x0-),(x1-),(xM-1-)]:R=XTX,ivi,i=0,1M-1,:i=1iXvi,i=0,1,M-1i,i,01M-1,,,861©1994-2007ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.:y=UTx,:^x=Uy[7]:R(x)=10lgx2x-^x2,R(x),x,Adaboost:ht(x|)=sign(R(x)-)3.3AdaboostAdaboost,,;,,,,,(,),Adaboost,(overfitting),,,Adaboost,Adaboost,:,HWt,HWt:Dt+1(xi)=Dt(xi)Zte-tht(xi)=yie-tht(xi)yi,Dt(xi)HWtetht(xi)yi,Dt(xi)HWt,Ztni=1Dt+1(xi)=1,HWtt,HWt,:HWt=ni=1Dt(xi)n,,DtHWt,,,,,,,:1):ht(x|)=sign(R(x)-),:{(x1,y1),(x2,y2),,(xn,yn)},xiX,X,yiY={-1,+1},-1,+12):Di(xi)=1n,i=1n3)Fort=1,2,......,T,DO:,i,i:ht(x)=sign(R(x)-i),i=argminiDt(xi)[ht(xi)yi]:t=ni=1Dt(xi)[ht(xi)yi],t=12ln1-tt:HWt=ni=1Dt(xi)n:Dt+1(xi)=Dt(xi)Zte-tht(xi)=yie-tht(xi)yi,Dt(xi)HWtetht(xi)yi,Dt(xi)HWt,Ztrni=1Dt+1(xi)=1,HWtt4)T,:H(x)=signnt=1tht(x),H(X)4ORL150,internet10050,3.12424,,Adaboost,H(X)MIT+CMU100,70,20,10:P2.4GMHzCPU,256M,WindowsXP,MATLAB6.5T50,AdaboostAdaboost,30,121,AdaboostAdaboost,,,Adaboost(253)961©1994-2007ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.:[1],,.()[M].:,1996.[2],,,,.[M].:,2001.[3]SPFortheringham.RogersonSpatialAnaysisandGIS[M].Lon2don:TaylorandFrancis,1994.[4]DepartmentofDefense,HighLevelArchitectureObjectModelTemplate,Version1.1[DB].12Feb,1997,downloadableatht2tp://hla.dmso.mil/.[5],.[J]..2000,(1):49-52.[](1965.9-),(),,,,;(1978.10-),(),,,,;(1968.7-),(),,,,,;(1981.9-),(),,,,(169)Adaboost,Adaboost1MIT+CMUMIT+CMUAdaboost9.6%17.1%Adaboost7.7%13.6%25Ada2boost,PCA,Adaboost,,AdaboostAdaboost,:[1]MHYang,DJKriegman,NAhuja.DetectingFacesinImages:ASurvey[J].IEEEPatternAnalysisMachineIntelligence,2002,24(1):34-58.[2].[J].,2002,,25(5):449-458.[3]PViola,MJones.RapidObjectDetectionUsingaBoostedCascadeofSimpleFeatures[C].In:ProceedingsofIEEEConferenceonComputerVisionandPatternRecognition,Kauai,Hawaii,USA,2001.[4]MKeams,LGValiant.LearningBooleanFormulaeorFactoring[R].TechnicalReportTR-1488,Cambridge,MA:HavardUni2versityAlikeComputationLaboratory,1988.[5]MKeams,LGValiant.CrytographicLimitationonLearingBooleanFormulaeandFiniteAutomate[C].In:Proceedingsofthe21stAnnualACMSymposiumonTheoryofComputing,NewYork,NY:ACMpress,1989:433-444.[6]YFreund,RESchapire.ADecision-TheoreticGeneralizationofOnlineLearingandanApplicationtoBoosting[J].JournalofComputerandSystemSciences,1997,55(1):119-139.[7],.()[M].:,2000,224-226.[8]MTurk,APentland.EigenfacesforRecognition[J].JournalofCognitiveNeuroscience,1991,3(1):71-86.[](1977-),(),,,,(1950-),(),,,:,,,,352©1994-2007ChinaAcademicJournalElectronicPublishingHouse.Allrightsreserved.