Eigenfaces for Recognition_kimo

EigenfacesforRecognitionMatthewTurkandAlexPentlandpresentedbyKimoJohnsonFaceRecognition•Faces–primaryfocusofattention–determineidentityandemotion•Humanability–speed–robusttochangesFaceRecognition•Computationalmodels–criminalidentification–securitysystems–human-computerinteraction•Goals–fast–reasonablysimple–accurateinconstrainedenvironmentsBackground•Individualfeatures–eyes,nose,mouth,headoutline–positionandsizerelationships•Disadvantages–multipleviews–fragileandcomplexEigenfaces•Theeigenfaceapproach–imagesarepointsinavectorspace–usePCAtoreducedimensionality–facespace•Sirovich&Kirby1987•Kirby&Sirovich1990–compareprojectionsontofacespacetorecognizefacesPCA•Principalcomponentanalysis–Xismxn•m:dimensionalityofimage•n:numberofimages–orthogonalchangeofvariable–maximizevarianceofprojectedsamples–eigenvectorsofcovariancematrixPCA•Optimization–WewanteigenvectorsofS(mxm)–Ifmismuchlargerthann,formT(nxn)EigenfaceRecognitionProcedure•Buildfacespace–PCA–chooseM’eigenfacesasabasisforfacespace•Projectimagevectorsontofacespace–nearestknownface(Euclideandistance)matches–thresholdsfordistancetofaceclassvs.distancetofacespace•infacespace,butnomatch•notinfacespaceExample:BuildFaceSpace40faces,112x92pixels=10,304pixelsExample:BuildFaceSpaceXis10,304x40,Tis40x40Example:BuildFaceSpaceFaceSpace=top8eigenfacesExample:RecognizeFacesSame40people,differentimagesExample:RecognizeFacesrecognize34/40=85%ExtensionsandOtherIssues•Extensions–locatinganddetectingfacesinimagesandvideo–recognizingnewfaces•Otherissues–eliminatingthebackground–scaleandorientationinvarianceConclusions•Facerecognitionsystem–fast–reasonablysimple–accurateinaconstrainedenvironment•Futurework–robustnesstochanges–learningnewfaces–eigenfacestodeterminegenderorfacialexpressionsPCAdetails•MaximizevarianceofprojectedsamplesPCAdetails•SolveusingLagrangemultipliers•Solutioniseigenvectorofcovariancematrix