用于人脸识别的深度神经网络(IJEM-V8-N1-6)

I.J.EngineeringandManufacturing,2018,1,63-71PublishedOnlineJanuary2018inMECS()DOI:10.5815/ijem.2018.01.06Availableonlineat:19May2017;Accepted:11September2017;Published:08January2018AbstractFacerecognition(FR),theprocessofidentifyingpeoplethroughfacialimages,hasnumerouspracticalapplicationsintheareaofbiometrics,informationsecurity,accesscontrol,lawenforcement,smartcardsandsurveillancesystem.ConvolutionalNeuralNetworks(CovNets),atypeofdeepnetworkshasbeenprovedtobesuccessfulforFR.Forreal-timesystems,somepreprocessingstepslikesamplingneedstobedonebeforeusingtoCovNets.Butthenalsocompleteimages(allthepixelvalues)arepassedasinputtoCovNetsandallthesteps(featureselection,featureextraction,training)areperformedbythenetwork.ThisisthereasonthatimplementingCovNetsaresometimescomplexandtimeconsuming.CovNetsareatthenascentstageandtheaccuraciesobtainedareveryhigh,sotheyhavealongwaytogo.Thepaperproposesanewwayofusingadeepneuralnetwork(anothertypeofdeepnetwork)forfacerecognition.Inthisapproach,insteadofprovidingrawpixelvaluesasinput,onlytheextractedfacialfeaturesareprovided.Thislowersthecomplexityofwhileprovidingtheaccuracyof97.05%onYalefacesdataset.IndexTerms:Facerecognition,haarcascade,deepneuralnetworks,convolutionalneuralnetworks,softmax.©2018PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheResearchAssociationofModernEducationandComputerScience.1.IntroductionFacerecognition(FR)systemidentifiesafacebymatchingitwiththefacialdatabase.Ithasgainedgreatprogressintherecentyearsduetoimprovementindesignandlearningoffeaturesandfacerecognitionmodels[1].Ashumanshaveanexceptionalabilitytorecognizepeopleirrespectiveoftheirage,lightingconditionsandvaryingexpressions.TheaimofresearchersistodesignanFRsystemwhichcanmatchorevensurpassthehumanrecognitionratewhichisnearly97.5%.Thetechniquesusedinbestfacialrecognitionsystemsmaydependontheapplicationofsystem.Facerecognitionsystemsmaybedividedintotwobroadcategories:Findapersonfromhisimageinalargedatabaseoffacialimages(eg.apolicedatabase).Thesesystems64DeepNeuralNetworkforHumanFaceRecognitionreturnsthedetailsofthepersonbeingsearchedfor.Oftenonlyoneimageisavailableperperson.Itisusuallynotnecessaryforrecognitiontobedoneinrealtime.Identifyapersoninrealtime.Theseareusedinsystemswhichallowaccesstoacertaingroupofpeopleanddenyaccesstoothers.Multipleimagesperpersonareoftenavailablefortrainingandrealtimerecognitionisrequired.Theproposedideaisforthesecondtypeofsystemswithvaryingfacialdetails,expressions,andangles.ItremainsanopenproblemtofindanidealfacialfeaturewhichisrobustforFRinunconstrainedenvironments[2].Theconventionalfacerecognitionpipelineconsistsoffourstages:facedetection,facealignment,facerepresentation(orfeatureextraction),andclassification[2].Theproposedmethodextractsfacialfeaturesfrominputimagesandfeedsthemtodeepneuralnetworksfortrainingandclassification(softmaxlayerisused).Thearchitectureofnetworkisveryflexibleandlayerscanbeaddedorremovedtogetbestresults.Inrecenttimestherearenumerouslibraries,functionsandplatformstocreateandmodifyanetwork.CovNetsareaspecializedkindofneuralnetworksforprocessingdatathathasaknown,grid-liketopology.Thesenetworkshavebeentremendouslysuccessfulinpracticalapplicationsthatincludetime-seriesdata,whichcanbethoughtofasa1Dgridtakingsamplesatregulartimeintervals,andimagedata,whichcanbethoughtofasa2Dgridofpixels.Convolutionalnetworksaresimplyneuralnetworksthatuseconvolutioninplaceofgeneralmatrixmultiplicationinatleastoneoftheirlayers.Thename“convolutionalneuralnetwork”indicatesthatthenetworkemploysamathematicaloperationcalledconvolution.Convolutionisaspecializedkindoflinearoperation[3].1.1.LiteratureWorkRecently,multipleCovNetsordeepCovNetshaveshowngoodresultsforfaceverification.AccordingtoYiSunet.al[1],existingmethodsgenerallyaddresstheproblemofFRintwosteps:featureextraction(designorlearnfeaturesfromeachindividualfaceimageseparatelytoacquireabetterrepresentation)andrecognition(calculatesimilarityscorebetweentwocomparedfacesusingfeaturerepresentationofeachface[1].Forfacerecognition(FR),manyapproacheshavebeenimplementedearlier,liketheuseofneuralnetworks[3,4,6],geometricalfeatures,Eigenfaces,templatematching,andgraphmatching.CovNetshasshownmanypromisingresultsforFR[2,4,5,7,8,9].Automaticfeatureextractionmethodusingratiosofdistances,presentedbyKanade[7]usedgeometricalfeaturesandreportedarecognitionratebetween45-75%withadatabaseof20people.Theapproacheslikeself-organizingmaps(SOM)andKarhunen-Loeve(KL)transformbothcanbeusedfordimensionalityreduction,fromwhichSOMprovedtobeanefficientalgorithm[3].PrincipalComponentAnalysis(PCA)hasalsobeensuccessfullyimplementedforsamepurpose[3,4].ThoughCovNetshaveshownpromisingresultsforFR,itremainsstillambiguoustodesignagoodCovNetarchitectureforaspecificclassificationtaskduetothelackof