基于改进特征点定位算法的人脸自动识别系统研究

上海交通大学硕士学位论文基于改进特征点定位算法的人脸自动识别系统研究姓名：费俊琳申请学位级别：硕士专业：生物医学工程指导教师：王志中2008010123ABSTRACT2ABSTRACTWiththedevelopmentofsocietyandtechnology,thenewdemandforcomputervisionandartificialvisionisincreasinglyurgent.Asthemostactivethemeinthecomputervisionandpatternrecognition,autohuman-facerecognitiontechnologyistheresearchfocuswithanextensiveapplicationforeground.Autofacerecognitionsystemiscomposedoffacedetection,facialfeaturelocation,facialrepresentationandfacematchingrecognition.Theresearchworkofthispaperfocusesonthefeaturedetectionandfacerepresentationtoimprovetherobustnessandefficiencyoffacerepresentationalgorithms.Inthisthesis,theimprovementofthetraditionaltemplatematchingalgorithm,basedonthefacepicturesdetectedbyAdboostalgorithm,foreyedetectioniscarriedoutfirstly.Theimprovementonaccuracyandrapidityofeyeslocatingarerevealedbythreeexperiments.ThenextmissionofthisthesisistheresearchontheGaborBunchFacialGraphforautofacerecognition.Themaininitiativeworksofthispaperareasfollows:Asthelimitationintheprecisionoftraditionaltemplatematching,thesisisproposesnewtemplatesdefinitionbyextendingeyetemplatestoeye-browmodel.Themodifiedtemplatesavoidthemisslocationcausedbythehighsimilaritybetweensomeeyesandsomebrows.Thethesisalsoproposestemplatematchingalgorithmforeyelocatingwithtemplateselectedbycorrelationcoefficient.Thealgorithmreducedtheredundancyofthetemplatesetbytakingthecorrelationcoefficienttoselecttheefficienttemplates,andthen,theman-madeerrorintemplateselectionisavoided.Thusthelocationprecisionandspeedareimprovedobviously.Thispapertakethebunchfacegraphasthetraintemplatefortrainset,andfortestsetatthesametime.SavetheBFGinthesystemmemory.Whenthesystemgetanewface,thesystemutilizetheBFGtopickupGaborvalueofthenewfaceandadditintototalsystemtrainsetABSTRACT3forrecognizethisnewface.Then,thesystemdoesnotneedupdatetheBFGfornewfacerecognition.Keywords:facerecognition,templatematching,eyeslocation,Gaborwavelet,elasticbunchgraphmatching2200812511.22Fig.1-1Facerecognitionprocess1.331.3.2451.41.51.5.16Fig.1-2SomedaceimagesintheORLfacedatabaseFig.1-3SomedaceimagesintheARfacedatabase7Fig.1-4SomedaceimagesintheYalefacedatabaseFig.1-5SomedaceimagesintheBioIDfacedatabase81.5.2Fig.1-6SomedaceimagesintheFERETfacedatabase1.6910Fig.2-1Faceautomaticrecognitionsystemproceed(facedetection)2.12.1.12.1.211122.2AdaBoost2.3AdaBoost13Fig.2-2Facedetectionsystemframe142.3.1Haar-LikeFig.2-3Haar-likefeatures(1,...,)Re()jiiiNfeaturectSumrw∈=∑(2-1)iwirRe()ictSumrirN15jfeatureirRe()ictSumri2.3.2IntegralImageI(,)xy'''',(,)(,)xxyySatxyixy≤≤=∑(2-2)(,)xy(,)xyFig.2-4Integralimagevalueofpoint(,)xy(,)xyFig.2-5Calculatetheintegralimagevalueofpoint(,)xyI(,)ixy16(,)SATxy(,)(,1)(1,)(1,1)(,)SATxySATxySATxySATxyixy=-+----+(2-3)(1,)(,1)(1,1)0SATySATxSAT-=-=--=Fig.2-6CalculatetherectangleD’sgreyintegralvalue3SATAB=+4SATABCD=+++1234DSATSATSATSAT=--+2.3.3Haar-Like()jjjjpfxpq()jhx=(2-4)17()jhxjqjp()jfxxFig.2-7SelectthebestHaar–likefeatures•••{}11(,),...,(,)mmSxyxy=Ηixc∈c{1,1}iy∈-+m11()Dim=1,...,im=181,...,tT=Ηhc12,,...,nXXXtD(,)()iiijlijitixXylWPxXylDi∈∧==∈==∑1l=±(2-5)jxX∀∈111()ln2jjWhxWee+-⎛⎞+=⎜⎟+⎝⎠1,...,jn=(2-6)e112jjjZWW+-=∑1,...,jn=(2-7)ZthZZminthZZ∈Η=(2-8)argminthhZ∈Η=(2-9)°1exp[()]()()itittyhxDiDiZ+-=1,...,im=(2-10)°tZ1tD+1()[()]TttHxsignhxb==-∑(2-11)192.4(2-11)Fig.2-8Cascadestructure20fdargtetF11F=1i=argitetFFibiffd1iiiFfF+←1ii←+Negf←1argitetFF+2.521Fig.2-9Multi-scalemagnifyfacedetectionandbasicdisplacementstep22Fig.2-10FacedetectedbysinglefacedetectionsystemFig.2-11Facepicsindatabasedetectedbyfacedetectionsystem232.624Fig.3-1FaceautomaticrecognitionsystemproceedEyesfeaturelocation3.13.1.1251()(,)NyIxixy==∑(3-1)Fig.3-2GrayprojectionfunctionforeyeslocationFig.3-3Eyesmisslocationcausedbydifferentreasons3.1.2Hough26Fig.3-4EyeslocationbasedonHoughtransferalgorithmf(,)ij(,)fij121000121xG⎛⎞⎜⎟=⎜⎟⎜⎟---⎝⎠101202101yG-⎛⎞⎜⎟=-⎜⎟⎜⎟-⎝⎠(3-2)xGyG(,)ijij(,,)eeijReiejR22()eeiirjj=+--(3-3)(,,)eeijR(,,)eeijR273.1.3Fig.3-5NegativeandpositiveeffectlocatitonofPixelsPwhenn=2283.1.4(,)cXr(,,,,)eXabcqeXabcqFig.3-6HumaneyetemplatewithVariableparametersvalleyE111valleyvalleyicirclecircleEEbArea==-∑(3-4)29valleyibedgeE21111[]2edgeedgeedgeiiiicircleparabolasEEbbLengthLength==-+∑∑(3-5)edgeibpeakE2311112iipeakpeakjijwindowwindowEEbArea=∈==-∑∑(3-6)peakjbiwindowcornerE241111[]2iivalleycornerjijwindowwindowEEbArea=∈==-∑∑(3-7)iwindow(3-8)15iciicirclecirclekEEgArea∈==∑(3-8)ig1k(3-9)30261iwiiwhiteswhitekEEgArea∈==-∑(3-9)ig2k(3-10)37ipiiparabolasparabolaskEEgLength∈==∑(3-10)ig3k3.23.2.139Fig.3-7Eyetemplates3111221,2221111(-)(-)()=(-)(-)cIIIIIIIIIIg∑∑∑(3-11)1I2I1I2Icg3.2.21/1212121(,)=L(,)()nppmPiiiMMMMmmg==-∑(3-12)21221212121(,)=L(,)()nmPiiiMMMMMMmmg===-=-∑(3-13)112212221122(-)(-)(,)=(-)(-)mMMMMMMMMMMg∑∑∑(3-14)1M2M321M1M2M2MmgmgFig.3-8FlowchartofTemplatesfiltering921×Fig.3-9Artificialselectionforeyetemplates33Fig.3-10MisseyeslocationbyeyetemplatesFig.3-11Eye-browtemplates34Fig.3-12Eyefeatureaaccuratelocationwitheye-browtemplates111(,)Pxy2exactarea=(21)step-2step=(,)Pxy2step=3step=Fig.3-13Roughlocationandaccur