A-face-recognition-system-based-on-BP-neural-netwo

AsystemofFaceRecognitionBasedonBPNeuralnetworkAbstractInthispaper,thewritertrytobuildafacerecognitionsystembasedonBackPropagationneuralnetworkinthispassage.Thesystemcanrecognizeapersonfaceimageofdifferentexpressions,differentshootingangles,differentilluminationintensity.Theaccuracycanreach98%.First,dopretreatmentto400facesamplesof40people.Eachpersonhastenimagesofdifferentfacialexpressions(smileornot),differentfacialdetails(withglassesornot).BymethodsofimagesegmentationandSVD(SingularValueDecomposition),wegetthefeaturematrixofthe400samples.What’smore,400onesaredividedintotwoparts,onefortraining,theotherfortesting.AndthenormalizedfeaturematrixpnareobtainedfortrainingastheinputoftheBPnet.Afterpretreatmentthefacesamplesanddividingtrainingsets,itistimetostarttoconstructtheBPneuralnetwork.Basedonbackpropagationalgorithmandtheneuralnetworktoolboxinmatlab,thenetisconstructed.BPneuralnetworkiscomposedofthreelayers:input,hiddenandoutput.Thenodenumberofhiddenlayerisabout110.Itisdecidedbyexperiencedformula.Thenwhattodoistotrainthenetatdifferenttimestomakeitstudyinordertoreachabestresult.KeyWords:Facerecognition,SVD,BPnetwork1.BackgroundintroductionInthisglobalizedworld,securityofapersonorcountryisamajorissue.Anannualbudgetforthecountryandtheirpeoplesecurityhasraisingdrasticallybyeverynation.Similareventscausedbynon-recognitionhaveledtofinancialanddeathloses.So,securityattheseeventsneedtobeenhancedtocontrolthevariousloses.AllthesesituationscanbeovercomedbyutilizingthetechnologysuchasartificialNeuralNetwork,Facerecognition,DigitalImageProcessingandsoon.[1]What’smore,academicworldandindustryworldhavepaidmoreandmoreattentiontofacerecognitionsincebiometricsidentification,HIC,objectdetectionhavedevelopedalottheseyears.Inaword,researchonfacerecognitionplayanimportantroleinboththeoryandtechnology,bothadvancedfieldandresidential.2.DevelopmentofresearchNowadays,manycountriesintheworldhaveconductedresearchonfacere-cognition.TheyaremainlyAmerica,theEuropes,Japan,andthefamousresearchinstitutionsincludeMedialab,AIlabofUS,theDepartmentofEngineeringinUniversityofCambridgeandsoon.Sincethe1990s,withthedevelopmentofhigh-speedcomputer,theresearchoffacerecognitionhasgotabreakup,steppinginthestageofautomaticdetection.Alotofuniversitieshavegotahugeprogress.Forexample,theAbdiandToolegroupofTexasatDallasinUS,mainlythelawsofhumanfaces.What’smore,professorBruceofStirlinguniversityandprofessorBurtonofGlasgowuniversity,theystudywhatarolehumanbrainsplayinfacerecognition..Inthefieldoffacerecognition,therearethefollowingdirections:1.Facerecognitionbasedongeometricfeatures.TherepresentativeistheBrunelliandPoggioofMIT.TheyimprovedintegralprojectionmethodtoextractcharacterizedbyEuclideandistanceof35-dimensionalfacialfeaturevectorsforpatternclassification2.Facerecognitionbasedontemplatematching.TherepresentiveisProfessorYuille.Heutilizestemplatetoextractthecontouroftheeyesandmouth,ChenandHuangfurthersuggestedextractingeyebrows,chinandnoseandotheruncertainshapewithactivecontourtemplate.3.FacerecognitionbasedonK-Ltransfer.RepresentativeisPentlandinMITlab.4.Facerecognitionbasedonneuralnetwork.PoggioposedbythePanelasHyperrBFneuralnetworkmethod,BuxtonandSussexUniversityinBritainHowellGroupproposedRBFnetworkidentificationmethods,et3.BPneuralnetworkModelbuildingandsolvingInthispaper,ItrytobuildasystembasedonBPneuralnetwork.Toworkbetter,Igoonthefollowingthreesteps:Facesamplesprepossessing,theconstructionandtrainingofBPneuralnetworkandthepicturesmatchingtestandanalysis.Thewholesystemcanbedescribedlikethis:4.Facesamplespretreatment4.1.FacesamplesIdownloaded400imagesof40people(eachpersonfortenpicturesofdifferentexpressionsanddifferentangles,withglassesornot)fromORL(OlivettiFig.1.ThesystemdiagramofthemodelResearchLaboratoryinCombridge,UK)database.Eachimageisa112*92grayscalematrixinfact.Someofthemaregivenoutasbelow:4.2.SegmentprocessingFirst,letthegrayscalematrixIbeforeachimage,.ThenIsegmentthematrixIintoseveralsmallmatrix.Accordingtoitsgrayscale,Itcanbedividedinto8,16,24,32,48,64,80.Anexamplesegmentofeightpartsasbelow:Fig.2.SomefacesamplesfromORLdatabaseThen,dosingularvaluedecompositionforiM(8......,,2,1i).4.3.SingularValueDecomposition（SVD）Thegrayscalematrixofimageisnotasquare.ThususingSVDtodecomposeit.AssumethatMisanmgraymatrix.Uisammmatrix,thecharacteristicsoftheTMMorthogonalvectoristherowofmatrixU.Visannmatrix,thecharacteristicsoftheMMTorthogonalvectoristherowofmatrixV.IfristherankofM,wecandoSVDlikethis:TVUM(1)MMMMTTandhavethesameeigenvalue:1,2,......,rForri1,letii,with1ii.Thenthenmmatrixiscomposedbysettingriiii1for,andzerootherwise.iisthesingularvalueoftheMmatrix.Thenwehave2TTTTMMUVVUUU(2)whereequation(2)isthedecompositionofsymmetricmatrices.Afterallthework,Wecangetthefeaturematrixfeatureofallsamples,itisamatrixof400*dim_num.5.TheconstructionandtrainingofBPneuralnetwork5.1.TheintroductionofBPnetworkBPnetworkisakindofartificialNeuralNetworkbasedonbackpropagationalgorithm.Infact,BPalgorithmisamethodtomonitorlearning.Itutilizesthemethod