利用图像颜色及其边缘直方图特征的SVM人脸检测

:2004207205:(A0210017).:,,1959,,,;,,1940,,,.SVM1,2,11(,230027)2(,362021)E2mail:[e-mail]:,.,.ROC,;,,.:;;;:TP391.41:A:100021220(2005)1222194206UsingColorandColorEdgeHistogramsforSVMBasedHumanFaceDetectionCHENDuan2sheng1,2,LIUZheng2kai11(DepartmentofElectricalEngineeringandInformationScience,UniversityofScience&TechnologyofChina,Hefei230027,China)2(DepartmentofComputerScience,HuaqiaoUniversity,Quanzhou362011,China)Abstract:Thispaperstudiesthehumanfacedetectionbysupportvectorclassifierusinghistogramsofcolor,coloredgeanditsorientation.Anoveledgeorientationcodingisproposed,itoutperformstheclassificationaccuracyofthetraditionaledgeorien2tationcodingwhentheyarecombinedrespectivelywithcolorhistogram.UsingROCevaluationinmulti_foldcrossvalidation,experimentsshowsthattheclassificationaccuracycanbesignificantlyimprovedwhenthecolorhistogramiscombinedwiththecoloredgehistograms.SVMcanmakeuseofthecolorbasedhistogramsforclassificationeffectively,andcandetectnon2deeprotatedhumanfaceincolorimageunderdifferentillumination,withdifferentexpressionsandpartialocclusion,whichshowsthatthecombinationofcoloranditsedgefeaturesiseffectiveandrobustforfacedetection.Keywords:coloredge;edgeorientation;supportvectormachine;receiveroperatingcharacteristic1,,,,.,,.[1].,,.,.,.2,.,.,[2].,.,,[325].,,.,,.,.,,;,.,2612200512MINI-MICROSYSTEMSVol126No.12Dec.2005,[6].,.ROC(ReceiverOper2atingCharacteristic).22.1.,,,,,[7],,[8],.f(x,y)RGB32,f=[f1,f2,f3]T,f1,f2,f33.D:D(x,y)=5f1ö5x5f1ö5y5f2ö5x5f2ö5y5f3ö5x5f3ö5y(1)xy(x,y).:DTD(x,y)=p(x,y)t(x,y)t(x,y)q(x,y)(2),.:p(x,y)=5f15x2+5f25x2+5f35x2t(x,y)=5f15x5f15y+5f25x5f25y+5f35x5f35yq(x,y)=5f15y2+5f25y2+5f35y2,(x,y)H:H=12tan-12t(p-q)(3)(x,y):G(x,y)=12[(p+q)+(p-q)cos2H+2tsin2H]1ö2(4)tan(A)=tan(AP),H0(3),H0+P2.,sin(A)=sin(A2P),(4),G(x,y),(4)[0,P).,DTDK,:G(x,y)=K(x,y)(5)K.1(a),(1)(2)(5)1(b)..(a)(b)(c)CEOM180(d)CEOM901,,18001,1(c),CEOM180(ColorEdgeOrientationMap,180).,0180.,CEOM1800180.,,18001090,1090180,1(d),CEOM90.4,,..,,.,,.2(d)(f).2.2,,.,,591212:SVM.[9].,RGB.(a)(b)(c)CEOM180(d)CEOM180(e)CEOM90(f)CEOM902Lenna4,.32,96,RGB3233.1,(2)2,,2436SVM.136(10),36,1(c)CEOM180,1(d)CEOM90.(5),64CGrads64.,SVM..,,[10].,,.,.,...33.1(x1,y1),,(xl,yl),xRn,y{+1,-1}yi[(wTxi-b]1,Pi=1,,l,,.,y{+1,-1}(x1,y1),,(xl,yl)xi.2::12w212wTw(6):yi[(wTxi-b],Pi=1,,L(7),(7).Ni0,i=1,,l,(7):yi[(wTxi)-b]1-Ni,Pi=1,,l(8),(8):12wTw+Cli=1Ni(9),C,C.C.,:W(A)=li=1Ai-12li,j=1AiAjyiyjxTixj(10)(0AiC),i=1,,l,li=1Aiyi=0.,(6)(10)..A0={A01,,A0l},:w0=li=1A0iyixi:f(x)=sign[svyiA0ixTix-b0](11)69122005xi,A0i,A0i0,.b0,b0=12[wT0x3(1)+wT0x3(-1)]x3(1)x3(-1)1-1.x:g(x)=li=1A0iyixTix-b0(12)xi,SVM,K(u,v)xTx.:5,K(u,v).,.K(u,v),.[11].(RBF),K(u,v)=exp{-u-v2öH},H0RBFRBF,()RBFHC.3.2ROC,.,SVM,,.3ROCROC.ROC[12].ROC,,.[1],TPR=P(predicated-trueûtrue)FPR=P(predicated-trueûfalse).(11);,(12),TPRFPR,(FPR,TPR)3ROC.ROC,.mnm,m.1,m21.m,m,,.m=n,,nn.5,5.,2502%[13].RBFHC,.,,[11].,,.,.44.1SVM,4..AR126791212:SVM4000160,,42.13CorelCD2500KodakPhotoCD,RGB3BP,,60160,,42.,320,.4.25ROCRGBöSVM,ROC5.CorrectnessP(predicated-trueûtrue).5(a)5(b)12RGB5SVMROCROC,3RGBROC,.5(a)5(b),SVM.5(c),,CEOM90CEOM180,5(a)(b),RGB,SVM,CEOM90CEOM180,.SVM(,),,RGBRGB,5(a)(b)(c)ROC.RGB,5(d)4,.,SVM,,,5(d).,89122005SVM.5,,,,.SVM,,,.,3,3264;6436,.,,..References:[1]Hjelma..sE,LowBK.Facedetection:Asurvey[J].ComputerVisionandImageUnderstanding,2001,83(3):2362274.[2]MoghaddamB,PentlandA.Probabilisticvisuallearningforob2jectrepresentation[J]IEEETrans.PatternAnalysisandMa2chineIntelligence,1997,19(7):6962720.[3]GagaudakisG,RosinP.IncorporatingshapeintohistogramsforCBIR[J].PatternRecognition,2002,35(1):81291.[4]WangJ,TanT.Anewfacedetectionmethodbasedonshapeinformation[J].PatternRecognitionLetter.2000,21(6,7):4632471.[5]JainA,VailayaA.Imageretrivalusingcolorandshape[J].PatternRecognition,199629(8):123321244.[6]ChenDuan2sheng,LiuZheng2kai.Amethodforautomaticde2tectionandcorrectionofhighlightedareaincolorfaceimage[J].JournalofSoftware,2003,14(11):190021906(inChinese).[7]GonzalezRC,WoodsRE.Digitalimageprocessing[M].2ndEd.Addison2WesleyPubCo,2002:3352338.[8]KoschanA.Acomparativestudyoncoloredgedetection[M].In:Proceedings2ndAsianConferenceonComputerVisionACCV’95,Singapore,528December1995,III,5742578.[9]MinC.ShinKyongI.ChangLeonidV.Tsap.Doescolorspacetransformationmakeanydifferenceonskindetection[].Wacv02[10]FuY,WangYW,WangWQetal.Content2baesdnaturalim2agevlassificationandretrievalusingSVM[J].CineseJouranlofComputers,2003,26(10):126121265.[11]VapnikVN.Thenatureofstatisticallearningtheory[M].NewYork:Springer2Verlag,2002:1002101,1762208.[12]TheodoridisS,koutroumbasKPatternrecognition[M].2ndEd,ElsevierScience,2003:1732174.[13]DudaRO,HartPE,StorkDG.Patternclassification[M].2ndEd.JoneWley&Son,2001:2592265,4832485.:[6],.[J]..2003,14(11):190021906.[10],,,.SVM[J].,2003,26(10):126121265.[11]VapnikVN.[M].:,2000:1002101,1762208.991212:SVM