当前位置:首页 > 机械/制造/汽车 > 汽车理论 > 图像自动识别外文文献原文
JBrazComputSoc(2012)18:229–235DOI10.1007/s13173-011-0049-zORIGINALPAPERSelectivelocaltexturefeaturesbasedfacerecognitionwithsinglesampleperclassK.JayaPriya·R.S.RajeshReceived:21April2010/Accepted:6November2011/Publishedonline:30November2011©TheBrazilianComputerSociety2011AbstractLocalappearance-basedmethodshavebeensuc-cessfullyappliedtofacerecognitionandachievedstate-of-the-artperformance.Inthispaperweproposealocalselec-tivefeatureextractionapproachbasedonGaborfiltersandtheLocalBinaryPattern(LBP)approachtofacerecogni-tion.AGaborfilterextractsthetexturalfeaturesfromthefaceimageandgeneratesthebinaryfacetemplateusingthosefeatures.Thebinaryfacetemplateactslikeamasktoextractthelocaltextureinformationofthefaceimageus-ingaLocalBinaryPatterntechnique.Thisselectivelocaltexturefeatureapproachusesthehistogram-basedmatch-ingforfacerecognition.Thismethodreducesthecomputa-tiontimeconsiderably.Thismethodalsoreducesthenum-berofLocalBinaryPatternsintohalfcomparedtotheexist-ingLBPmethod.Thisproposedapproachreducesthecom-putationtimefortheFERETdatasetby45%.Experimentsonwell-knownfacedatabasessuchasFERET,Yale,IndianFacesandORLshowthatthisapproachobtainsconsistentandpromisingresultsinthescenarioofonetrainingsampleperpersonwithsignificantfacialvariation.KeywordsFacerecognition·LocalBinaryPattern·Gaborfilter·Binaryfacetemplate·Expressioninvariantfacerecognition·SingleSampleProblemK.JayaPriya()ResearchScholar,DepartmentofComputerScience,MotherTeresaWomen’sUniversityKodaikanal,Kodaikanal,624102,Indiae-mail:kjp.jayapriya@yahoo.comR.S.RajeshDepartmentofComputerScienceandEngineering,ManonmaniamSundaranarUniversity,Tirunelveli,627012,Tamilnadu,Indiae-mail:rsrajesh_cse@msuiv.ac.in1IntroductionInrecentyears,facerecognitionhasreceivedmuchattentioninmanyareassuchasentertainment,informationsecurity,lawenforcement,andsurveillance[36].Mostoftheface-recognitionmethodssuchaseigenfaces[26],Fisherfaces[4]andLaplacianfaces[9],nearestfeatureline-basedsub-spaceanalysis,neuralnetworks[14,25],elasticbunchgraphmatching[30]andkernelmethods[33]wereinitiallydevel-opedwithfaceimagescollectedunderrelativelywellcon-trolledconditions,andinpracticetheyhavedifficultyindealingwiththerangeofvariationoftheappearancethatcommonlyoccursinunconstrainednaturalimagesduetoillumination,pose,facialexpression,ageingandpartialoc-clusions.Anothermostchallengingproblemforfacerecog-nitionistheso-calledSingleSampleProblem(SSP),wherethetrainingprocessusesasinglesamplepersubject.Insomespecificscenarios,suchaslawenforcement,passportoridentificationcardverificationetc.,onlyoneimageperpersonisavailablefortraining.Someface-recognitional-gorithms[5,6,18,24,25]wereproposedtosolvetheone-sampleprobleminvariousprocessmodes.Face-recognitionmethodsaregenerallydividedintotwocategories:holisticmatchingmethodsandlocalmatchingmethods.Theholisticmatchingapproachesusethewholefacere-gionasinputtotheface-recognitionsystem.TheprincipleofholisticmethodsistobuildasubspaceusingPrincipalCom-ponentAnalysis(PCA)[26],LinearDiscriminantAnaly-sis(LDA)[4,7,31]orIndependentComponentAnalysis(ICA)[3].Thefaceimagesarethenprojectedandcomparedinalow-dimensionalsubspacetoavoidthecurseofdimen-sionality.WangandTang[27]haveunifiedPCA,LDAandBayesianmethodsintothesameframeworkandpresentamethodtofindtheoptimalconfigurationforLDA.Inordertohandlethenonlinearityinfacefeaturespace,nonlinear230JBrazComputSoc(2012)18:229–235kerneltechniquessuchaskernelPCA[23],kernelLDA[18]etc.arealsointroduced.Comparedwithholisticmethods,localmethodsaremoresuitableforhandlingtheone-sampleproblemduetothefollowingobservations:Firstly,inlocalmethods,thelow-dimensionallocalfeaturevectorsrepresenttheoriginalfaceratherthanonesinglefullhigh-dimensionalvector.Thusthe“curseofdimensionality”isalleviatedfromthebegin-ning.Secondly,localmethodsoffermoreflexibilitytorec-ognizeafacebasedonitsparts;thusthecommonandclass-specificfeaturesareeasilyidentified.Thirdly,differentfa-cialfeaturesimprovetheclassifiersdiversity,whichishelp-fulforfaceidentification.Thelocalmatchingapproacheshaveshownsomepromisingresultsinfacerecognition[1,2,8,12,13,15–17].Thesemethodsfirstextractseveralfa-cialfeaturesandthenmakeacomparisononthebasisoflocalstatisticsforrecognition.Thecomparisonoflocalap-proacheswithglobalapproachesshowsthatthelocalsys-temoutperformedtheglobalsystemwith60%[10].Thereexistseverallocalappearance-basedmethodsforextractingthemostusefulfeaturesfromfaceimagestoaddressfacerecognition.TheLocalBinaryPattern(LBP)method[19]wasorig-inallyproposedasanimagetexturedescriptor[20],butitalsoappliedonface-recognitionapplication[1].Facerecog-nitionusingtheLBPmethod[2]providesverygoodresults,bothintermsofspeedanddiscriminationperformance.MorerecentworkonLBP[28],theHeatKernelLocalBi-naryPattern(HKLBP)descriptor,extractsmultiscaleHeatKernelStructuralInformation(HKSI)matricestocapturetheintrinsicstructuralinformationofthefaceappearance.Then,theLocalBinaryPatternanalysisonHKSImatricesprovidestheHKLBPdescriptorfortherepresentationoftheface.ThefeatureextractionwithLBPisastraightfor-ward(real-time)proc
本文标题:图像自动识别外文文献原文
链接地址:https://www.777doc.com/doc-4515227 .html