资源描述:
《FACIAL SIMILARITY ACROSS AGE,DISGUISE,ILLUMINATION AND POSE》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、FACIALSIMILARITYACROSSAGE,DISGUISE,ILLUMINATIONANDPOSENarayananRamanathan,RamaChellappaAmitK.RoyChowdhury∗CenterforAutomationResearchandDept.ofElectricalEngineeringDept.ofElectrical&ComputerEngineeringUniversityofCalifornia,UniversityofMaryland,CollegePark,MD2
2、0742Riverside,CA92521ABSTRACTlinearlyprojectingtheimagespaceontoalowdimensionalfea-turespacethatspansthesignificantvariationsamongknownfaceIllumination,posevariations,disguises,agingeffectsandex-images.Thesignificantfeaturesarenothingbuttheeigenvectorspressionva
3、riationsaresomeofthekeyfactorsthataffecttheper-ofthesetoffaces.TheEigenfacesmethodyieldsprojectiondirec-formanceoffacerecognitionsystems.Facerecognitionsystemstionsthatmaximizethetotalscatteracrossallclasses,i.e.,acrossallhavealwaysbeenstudiedfromarecognitionp
4、erspective.Ourimagesofallfaces.Thusunwantedvariationsduetoilluminationemphasisisonderivingameasureofsimilaritybetweenfaces.changesareretained.TheFischerfacesmethodselectstheprojec-Thesimilaritymeasureprovidesinsightsintotheroleeachofthetiondirectionbymaximizin
5、gtheratiooftheinter-classscatterma-abovementionedvariationsplayinaffectingtheperformanceoftrixtotheintra-classscattermatrix.ThustheFischerfacesmethodfacerecognitionsystems.Intheprocessofcomputingthesimi-isbetterequippedtohandleilluminationvariationsthantheEige
6、n-laritymeasurebetweenfaces,wesuggestaframeworktocompen-facesmethod.TheSubspaceLDAmethodprojectsfaceimagessateforposevariationsandintroducethenotionof`Half-faces'fromtheoriginalvectorspacetoafacesubspaceusingPrincipaltocircumventtheproblemofnon-uniformillumina
7、tion.WeusedComponentAnalysisandthenusesLinearDiscriminantAnalysisthesimilaritymeasuretoretrievesimilarfacesfromadatabasetoobtainalinearclassifierinthesubspace.SubspaceLDAandcontainingmultipleimagesofindividuals.Moreover,wedevisedFischerfaces[4,5]reportbetterrec
8、ognitionresultsthantheEigen-experimentstostudytheeffectageplaysinaffectingfacialsimi-facesmethod.HoweverboththeSubspaceLDAmethodandthelarity.Inconclusion,thesimilaritymeasurehelpsi