学习基于人脸识别巴氏距离的BP神经网络进行误差函数(IJIGSP-V4-N8-2)

I.J.Image,GraphicsandSignalProcessing,2012,8,8-14PublishedOnlineAugust2012inMECS()DOI:10.5815/ijigsp.2012.08.02Copyright©2012MECSI.J.Image,GraphicsandSignalProcessing,2012,8,8-14LearningaBackpropagationNeuralNetworkWithErrorFunctionBasedonBhattacharyyaDistanceforFaceRecognitionNaouarBelghinia,ArsalaneZarghilib,JamalKharroubic,AichaMajdadSidiMohamedBenAbdellahUniversity,FSTF,[e-mail],[e-mail],[e-mail],[e-mail]—Inthispaper,acolorfacerecognitionsystemisdevelopedtoidentifyhumanfacesusingBackpropagationneuralnetwork.ThearchitectureweadoptisAll-Class-in-One-Network,wherealltheclassesareplacedinasinglenetwork.ToacceleratethelearningprocessweproposetheuseofBhattacharyyadistanceastotalerrortotrainthenetwork.IntheexperimentalsectionwecomparehowthealgorithmconvergeusingthemeansquareerrorandtheBhattacharyyadistance.Experimentalresultsindicatedthattheimagefacescanberecognizedbytheproposedsystemeffectivelyandswiftly.IndexTerms—Backpropagation,NeuralNetwork,Facerecognition,Errorfunction,BhattacharyyadistanceI.INTRODUCTIONFacerecognitioncanbedefinedastheabilityofasystemtoclassifyordescribeahumanface.Themotivationforsuchsystemistoenablecomputerstodothingslikehumandoitsotoapplycomputerstosolveproblemsthatinvolveanalysisandclassification.Researchinthisareahasbeenconductedformorethan30years;asaresult,thecurrentstatusofthefacerecognitiontechnologyiswelladvanced.FacerecognitionhasreceivedthisgreatdealofattentionbecauseofitsapplicationsinvariousdomainslikeSecurity,identityverification,videosurveillance,Criminaljusticesystemsandforensic,Multi-mediaenvironments[12].Manyresearchareaaffectthefieldoffacerecognition:patternrecognition,computervision,neuralnetworks,machinelearning,etc…Neuralnetworkshavebeenwidelyusedforapplicationsrelatedtofacerecognition.OneofthefirstneuralnetworkstechniquesusedforfacerecognitionisasinglelayeradaptivenetworkcalledWISARD[13].Manydifferentmethodsbasedonneuralnetworkhavebeenproposedsincethenandthemajorofthemuseneuralnetworksforclassification.In[16]RadialBasisneuralnetworkwasusedtodetectfrontalviewsoffaces,curvelettransformandLinearDiscriminantAnalysis(LDA)wereusedtoextractfeaturesfromfacialimages,andradialbasisfunctionnetwork(RBFN)wasusedtoclassifythefacialimagesbasedonfeaturestakenfromORLdatabase.In[14]thefocuswastoinvestigatethedimensionalityreductionofferedbyrandomprojection(RP)andperformafacerecognitionsystemusingbackpropagationneuralnetwork.Experimentsshowthatprojectingthedataontoarandomlower-dimensionalsubspaceyieldsresultsandgiveanacceptablefacerecognitionrate.Bhattacharjeeetal.developedin2009afacerecognitionsystemusingafuzzymultilayerperceptronusingbackpropagation[15].Thewayinconstructingtheneuralnetworkstructureiscrucialforsuccessfulrecognition.Ingeneral,neuralnetworksaretrainedtominimizeasquaredoutputerrorwhichisequivalenttominimizingtheEuclidiannormofthedifferencebetweenthetargetandpredictionvectors.Themajorlimitationsofthisalgorithmaretheexistenceoftemporary,localminimaresultingfromthesaturationbehavioroftheactivationfunction,andtheslowratesofconvergence.manyresearchershavebedonetoovercometheseproblems.Inthiscontext,anumberofapproacheshavebeenimplementedtoimprovetheconvergencespeed.Therearebasicallyonselectionofdynamicvariationoflearningrateandmomentum,selectionofbetteractivationfunctionandbettercostfunction.Somerelatedresearchesarepresentedin[3].Untilrecentyearsgray-scaledimageswereusedtoreduceprocessingcost[4][5].Nowadays,itisdemonstratedthatcolourinformationmakescontributionandenhancesrobustnessinfacerecognition[6].OuraiminthispaperistointroducetheuseofcolorinformationwithminimalprocessingcostusingtheBackPropagationalgorithm,andtoacceleratethetrainingprocedureweuseBhattacharyyadistanceinsteadofthetraditionalmeansquareerror(MSE)basedontheEuclidiandistancetoseekstheglobalminimaontheerrorsurface.Theremainderofthepaperisorganizedasfollows:LearningaBackpropagationNeuralNetworkWithErrorFunctionBasedon9BhattacharyyaDistanceforFaceRecognitionCopyright©2012MECSI.J.Image,GraphicsandSignalProcessing,2012,8,8-14Section(2)givesanoverviewaboutthebackpropagationneuralnetwork.Section(3)presentstheBhattacharyyadistanceforsimilaritymeasurement.Insection(4),wepresentourproposalsolutionforfacerecognition.Section(5)givestheexperimentalresults.Finally,Section(6)givesaconclusion.II.THEBACKPROPAGATIONNEURALNETWORKBackpropagationisamulti-layerfeedforward,supervisedlearningnetwork.Thenetworkrepresentsachainoffunctioncompositionswhichtransformaninputtoanoutputvector.Itlookstominimizetheerrorfunctionusingthemethodofgradientdescent.Thelearningproblemconsistsoffindingtheoptimalcombinationofweightssothatthenetworkfunctionapproximatesagivenfunctionascloselyaspossible.Inotherword,gradientdescentisusedtoupdateweightstominimizethesquarederrorbetweenthenetworkoutputvaluesandthetargetoutputvalues;Theupdaterulesarederivedbytakingthepartialderivativeoftheerrorfunctionwithrespecttotheweightstodetermineeachweight’scontributiontotheerror