

PatternRecognition47(2014)1547–1558
ContentslistsavailableatScienceDirect
PatternRecognition
journalhomepage:doc.guandang.net
Doublelinearregressionsforsinglelabeledimageperpersonfacerecognition
FeiYina,n,L.C.Jiaoa,FanhuaShangb,LinXionga,ShashaMaoa
KeyLaboratoryofIntelligentPerceptionandImageUnderstandingofMinistryofEducationofChina,XidianUniversity,Mailbox224,No.2SouthTaiBaiRoad,Xi’an710071,PRChinab
DepartmentofElectricalandComputerEngineering,DukeUniversity,Durham,NC27708,USA
a
articleinfo
Articlehistory:
Received12March2013Receivedinrevisedform10September2013
Accepted19September2013Availableonline9October2013
Keywords:
Semi-superviseddimensionalityreductionLabelpropagationSparserepresentationLinearregressions
LineardiscriminantanalysisFacerecognition
abstract
Recentlytheunderlyingsparserepresentationstructureinhighdimensionaldatahasreceivedconsiderableattentioninpatternrecognitionandcomputervision.Inthispaper,weproposeanovelsemi-superviseddimensionalityreduction(SDR)method,namedDoubleLinearRegressions(DLR),totackletheSingleLabeledImageperPerson(SLIP)facerecognitionproblem.DLRsimultaneouslyseeksthebestdiscriminatingsubspaceandpreservesthesparserepresentationstructure.Speci cally,aSubspaceAssumptionbasedLabelPropagation(SALP)method,whichisaccomplishedusingLinearRegressions(LR),is rstpresentedtopropagatethelabelinformationtotheunlabeleddata.Then,basedonthepropagatedlabeleddataset,asparserepresentationregularizationtermisconstructedviaLinearRegressions(LR).Finally,DLRtakesintoaccountboththediscriminatingef ciencyandthesparserepresentationstructurebyusingthelearnedsparserepresentationregularizationtermasaregularizationtermofLinearDiscriminantAnalysis(LDA).Theextensiveandencouragingexperimentalresultsonthreepubliclyavailablefacedatabases(CMUPIE,ExtendedYaleBandAR)demonstratetheeffectivenessoftheproposedmethod.
&2013ElsevierLtd.Allrightsreserved.
1.Introduction
Inmany eldsofscienti cresearchsuchasfacerecognition[1],bioinformatics[2],andinformationretrieval[3],thedataareusuallypresentedinaveryhighdimensionalform.Thismaketheresearchersconfrontwiththeproblemof“thecurseofdimensionality”[4],whichlimitstheapplicationofmanypracticaltechnologiesduetotheheavycomputationalcostinhighdimen-sionalspace,anddeterioratestheperformanceofmodelestima-tionwhenthenumberoftrainingsamplesaresmallcomparedtothenumberoffeatures.Inpractice,dimensionalityreductionhasbeenemployedasaneffectivewaytodealwith“thecurseofdimensionality”.Inthepastyears,avarietyofdimensionalityreductionmethodshavebeenproposed[5–10].
Accordingtothegeometricstructureconsidered,theexistingdimensionalityreductionmethodscanbecategorizedintothreetypes:globalstructurebasedmethods,localneighborhoodstructurebasedmethods,andtherecentlyproposedsparserepresentationstructure[11,12]basedmethods.Twoclassicaldimensionalityreductionmeth-odsPrincipleComponentAnalysis(PCA)[13]andLinearDiscriminantAnalysis(LDA)[14]belongtoglobalstructurebasedmethods.Inthe eldoffacerecognition,theyareknownas“Eigenfaces”[15]and“Fisherfaces”[16].Twopopularlocalneighborhoodstructurebased
n
Correspondingauthor.Tel.:þ862988202661;fax:þ862988201023.E-mailaddress:yinfei701@163.com(F.Yin).
methodsareLocalityPreservingProjections(LPP)[17]andNeighbor-hoodPreservingEmbedding(NPE)[18].LPPandNPEarenamed“Laplacianfaces”[19]and“NPEfaces”[18]infacerecognition.TherepresentativesparserepresentationstructurebasedmethodsincludeSparsityPreservingProjections(SPP)[20],SparsityPreservingDiscri-minantAnalysis(SPDA)[21]andFastFisherSparsityPreservingProjections(FFSPP)[22].Theyhavealsobeensuccessfullyappliedtofacerecognition.Inordertodealwiththenonlinearstructureindata,mostoftheabovelineardimensionalityreductionmethodshavebeenextendedtotheirkernelizedversionswhichperforminReproducingKernelHilbertSpace(RKHS)[23].KernelPCA(KPCA)[24]andKernelLDA(KLDA)[25]arethenonlineardimensionalityreductionmethodscorrespondingtoPCAandLDA.KernelLPP(KLPP)[17,26]andKernelNPE(KNPE)[27]arethekernelizedversionsofLPPandNPE.ThenonlinearversionofSPDAisKernelSPDA[21].
Oneofthemajorchallengestoappearance-basedfacerecogni-tionisrecognitionfromasingletrainingimage[28,29].Thisproblemiscalled“onesampleperperson”problem:givenastoreddatabaseoffaces,thegoalistoidentifyapersonfromthedatabaselaterintimeinanydifferentandunpredictableposes,lighting,etc.fromjustoneimageperperson[28].Undermanypracticalscenarios,suchaslawenforcement,driverlicenseandpassportcardidenti cation,inwhichthereisusuallyonlyonelabeledsampleperpersonavailable,theclassicalappearance-basedmeth-odsincludingEigenfacesandFisherfacessufferbigperformancedroportendtofailtowork.LDAfailstoworksincethewithin-classscattermatrixdegeneratestoazeromatrixwhenonlyone
0031-3203/$-seefrontmatter&2013ElsevierLtd.Allrightsreserved.

doc.guandang.net
什么是Double Double技巧最近被各个论坛传的很神奇,咱们今天来看看,到底什么是Double 说来很奇怪,我第一次在国外的文章中接触到的Double实际上写的是Doubling,后来...
3 安装 Double-Take Backup 软件 ... 4 CDP 服务器安装 ......
130322_不要再说double confirm了!_人力资源管理_经管营销_专业资料。不要再说"double confirm"了! ? (来源.世界公民文化中心) 有一次我听到学生和老师的对话,...
Nutriway Double X_调查/报告_表格/模板_应用文书。Nutriway Double XNutriway? Double X ?崔莱 Double X 综合营养片 由於现代人的生活方式、食用高加工食品、...
I called to confirm again. 记住了,不要再说 double confirm,用 double check 或 reconfirm/confirm again/ for second confirmation 都可以。 ...
double可以是名词, 动词(vi&vt), 形容词, 副词, 限定词; double做限定词时用法与twice相同 double跟在动词后面做副词, 用法同twice (see double) --> double...
matlab中的double_数学_自然科学_专业资料。matlab 默认的数据类型是 double(64 位) 。 double 是双精度浮点数,保留 16 位有效数字 。 双精度到底是几位,与...
float和double的存储方式_计算机软件及应用_IT/计算机_专业资料。C/C++的浮点数据类型有 float 和 double 两种。 类型 float 大小为 4 字节,即 32 位,内存中...
实时备份及高可用容灾软件中铁信息工程集团 茅建中 Vision Solutions Double Take系列 VISON Solutions IBM 合作伙伴 IBM PureFlex 认证应用之一 支持AIX 支持IBM I ...
Double() 功能将字符串或 Blob 值转换成 Double 类型的值。 语法 Double(stringorblob) 参数 stringorblob:string 类型变量或 Blob 类型变量。Blob 类型变量时,把...
热门文档