一维小波转换在生物辨识的应用

Applying1-DimensionWaveletTransformtoBiometricRecognition***840TEL:[phone]E-mail:*[e-mail],**[e-mail]~chen/—(LPCC)(Fractal)(WTFT)[1](PNN)ORL;AbstractThispaperproposesablockdiagramappliedinthemainproblemoftwobiometricrecognitions—speakerrecognitionandfacerecognition.Firstly,tospeakerrecognition,waveletdecomposesspeechsignalintohighandlowfrequencycoefficients,andcombinessomemethodsincludingLPCC,fractal,WTFTandPCAforfeatureextraction.ThefeaturesareputinPNNforvoiceprintmatching.Theproposedmethodimprovesperformanceandefficiencyinspeakerrecognitionsystem.Secondly,tofacerecognition,traditionalfeatureextractionisto2-Dormorefaceimageresultinginlowefficiency,butthispaperuseshorizontalprojectionof2-Dimagetoobtainaccumulatedenergyprofilesignal.Thenweadopt1-Ddiscretewavelettransformtoextractlowfrequencycoefficientsasfeaturevectoroffacialimage.Forfaceidentificationandmatchingapplicationmodes,weproceedasetofexperiments.ThefacialimagesaresampledfromORLdatabase.Ourexperimentsrevealthattheproposedmethodpossessesexcellentrecognitionperformanceandefficiency.Itisadvantageoustorealizeafacialrecognitionsysteminahardware-friendlyresource-constrainedembeddedenvironment.Atpresent,singlebiometricsfeaturesystemhardlycopeswiththerequirementofsocial.Infuture,integratedtwoormorebiometricsfeaturesystemimprovesrecognitionrateandsafe.Keyword:speakerrecognition;facerecognition;featureextraction;wavelettransform1.(BiometricsRecognition)(LPCC)[8](PCA)[8,9](Fractal)(facedetection)(featureextraction)[1,2,3,4,5,6,7][7]()(PrincipalComponentAnalysis,PCA)[2,6](FaceIdentification)1(a)(Matchingorverification)/1(b)/(a)(b)1.(a);(b)22.122.(LPCC)[8](PCA)[8,9]2.2(LPCC)(LPC)LPCLPCCLPCC(Robust)(Reliability)LPCLPCCLPCCLPCLPC(Autocorrelation)()()[]()1,,0+=+=∑∞-∞=pkkRknxnxkRnL()nx()kRpLPCLevinsion-Durbin[11]LPC()()00RE=pifor:1=()()()()()1111--=-=--=∑iiiiijijiEkjiRiRkaa1:1-=ijfor()()()11----=ijiiijijkaaaend()()()121--=iiiEkEendLPC()piatscoefficienLPCpii≤≤==1aLPCLPCCmCpmacmkacmkkmkmm≤≤⎟⎠⎞⎜⎝⎛+=∑-=-111LPCCLPC2.3(PCA)X(Row)M()()'MXMXE--=EPPA′=PE[9]2.4(BenoitB.Mandelbrot)[12](IterativeFunctionSystem,IFS)jW(FixedPoint)IFS(attractor)⎥⎦⎤⎢⎣⎡+⎥⎦⎤⎢⎣⎡⎥⎦⎤⎢⎣⎡=⎥⎦⎤⎢⎣⎡iiiiiifeyxdcayxW0⎥⎦⎤⎢⎣⎡=⎥⎦⎤⎢⎣⎡--1100iiiFxFxW⎥⎦⎤⎢⎣⎡=⎥⎦⎤⎢⎣⎡iiNNiFxFxWxiFiiNiNiiiiiiiNiiiiFfFdxcFfFdxcxexaxexa=++=++=+=+--10010iaicidieifIFSixiaieid(ScalingFactor)()2minmaxmaxiiiFFFd-=IFSicidif(Covariancematrix)R*[]iiiifdcx=*[]TNxxxA**2*1*L=∑==NiijjxNm1*1()∑=-=NijijjmxNS12*21jm2jSj()jjijijSmxx-=*[]tNxxxAL21=ARAANRt1=3.3.1(Eigenface)X(Column)M()()'MXMXE--=EPPA′=PE[2](LDA)[3]CMcmcmx(Within-ClassScatterMatrix)()()∑∑==--=CcMmTccmccmwxxxxSx11(1)cxc(Between-ClassScatterMatrix)()()∑=--=CcTccbxxxxSx1(2)xAcmTcmxAy=(3)FyywbSSF=(4)Accx∑==MmcmcxMx11(5)x∑∑===CcMmcmxCMx111(6)ccycTMmcmcxAyMy==∑=11(7)xAyCMyTCcMmcm==∑∑==111(8)(7)(8)ASASxywTw=(9)ASASxybTb=(10)(4)ASAASAFyywTbT=(11)(11)ASAASAAyywTbTAoptmaxarg=(12)(12)()()ASAASAAASSxxxxbTwTbw11--=(13)(13)AxxbwSS1-[3]3.2(dimensionreduction)2(a)112x92⎥⎥⎥⎦⎤⎢⎢⎢⎣⎡=92112111292111xxxxxxxxLMOMLX(14)y∑==921jijixy,921,1121ji(15)⎥⎥⎥⎦⎤⎢⎢⎢⎣⎡=1121yyMY(16)112x92112x13(b)3(a)3.(a)(b)44.1(WaveletTransform,WT)[6,13,14,15]4s[n]HLD(DetailSignal)A(ApproximateSignal)3(a)H~L~y[n](17)(18)(17)(18)s[n]D[n]A[n][][][]HsDknHksnDk,=⇔-⋅=∑∞-∞=(17)[][][]LsAknLksnAk,=⇔-⋅=∑∞-∞=(18)(Multi-resolutionanalysis,MRA)H~L~(a)LHH][nsLLHA2D1D3D(b)4.(a)(b)4.2l(WTFT)1,,2,1112,+==∑=NiwnvinjjiiiLinNiv{}tivvvV,,,21L=lLPCC,PCAFractal(PNN)4.33.2535.5D.F.specht1988(ProbabilisticNeuralNetwork,PNN)[11]65.11.2.3.2s6P01()()'YXYXE-⋅′-=2dEeP-=6.5.2(Identification)XYXPNN2dEeP-=(19)()()'YXYXE-⋅′-=(20)NPNNNP(20)P5.3(Matching)(verify)(ContingenceTable)HH()()(FalseRejectionRate,FRR)H(FalseAcceptanceRate,FAR)FARFRR7FARFRR(EqualErrorRate,ERR)(recognitionrate)[16]6l43542030”xuziheng”(wav)6911kHZ83(32000-4000)520AMD750Hz378MSdramWindowsXPProfessionalMatlab6.1l(LPCC)(PCA)(WTFT)(Fractal)1011FractalPCALPCCWTFT0.710.670.950.97361013248(ms/sample)1002291154(ms/sample)17070561221lFractalFractalLPCC12LPCCPCAPCA22WT+FractalWT+LPCCWT+PCA0.8950.9950.61368410(ms/sample)776131(ms/sample)17036070lEER343FractalLPCCPCAWTFTEER0.19170.11670.40.054WT+FractalWT+LPCCWT+PCAEER0.100.38LPCCPCA7,[2,3]ORL[17]4010112x92AMD750Hz384MSdramWindowsXPMatlab6.17.1(Identification)10561051(M1)db22(M2):db2(Eigenface)[2,4,5]3(M3)db2(LDA)[3,6,15]5.M1M2M3%98.596.595.5%95.691.493.7163030(s)9.4331.8227.17(s)0.050.090.095Matlab6.114009.43231/30.0523216LDA307.2(Matching)EER66.M1M2M3EER0.0240.0550.0447FARFRRERREER=0.024Daubechies2(EigenfaceLDA)7(EER)8LPCCPCA[1]R.Chellappa,C.L.Wilson,andS.Sirohey,“Humanandmachinerecognitionoffaces:Asurvey,”Proc.IEEE,vol.83,pp.705–740,1995.[2]BaiLi*,YihuiLiu“Wheneigenfacesarecombinedwithwavelets”,Knowledge-BasedSystems,Vol.15(2002),pp343-347.[3]“”2000[4]M.A.TurkandA.P.Pentland,“Eigenfacesforrecognition,”J.CognitiveNecurosci.,vol.3,pp.71-86,1991.[5]T.Phiasai,S.Arunrungrusmi,andK.Chamnongthai,”FacerecognitionsystemwithPCAandmomentinvariantmethod,”inProc.IEEEInt.Symp.CircuitsSyst.,2001,pp.III65III68.[6]B.-L.ZhangandY