Large Margin Coupled Mapping for Low Resolution Face Recognition

《Large Margin Coupled Mapping for Low Resolution Face Recognition》由會員上傳分享，免費在線閱讀，更多相關內(nèi)容在學術論文-天天文庫。

1、LargeMarginCoupledMappingforLowResolutionFaceRecognition1(&)1,3,412JiaqiZhang,ZhenhuaGuo,XiuLi,andYoubinChen1GraduateSchoolatShenzhen,TsinghuaUniversity,Shenzhen,Chinazhang-jq13@mails.tsinghua.edu.cn2HuazhongUniversityofScienceandTechnology,Wuhan,China3KeyLaboratoryofMeasurementandControlofCo

2、mplexSystemsofEngineering,MinistryofEducation,SoutheastUniversity,Nanjing,China4KeyLaboratoryofIntelligentPerceptionandSystemsforHigh-DimensionalInformation,MinistryofEducation,NanjingUniversityofScienceandTechnology,Nanjing,ChinaAbstract.Traditionalfacerecognitionalgorithmscanachievesigni?ca

3、ntperformanceunderwell-controlledenvironments.However,thesealgorithmsperformpoorlywhentheresolutionofthefaceimagesvaries.Atwo-stepframeworkisproposedtosolvetheresolutionproblemthroughadoptingsuper-resolution(SR)andperformingfacerecognitiononthesuper-resolvedfaceimages.However,suchmethodusuall

4、yhaspoorperformanceonrecog-nitiontasksasSRfocusesmoreonvisualenhancement,ratherthanclassi?cationaccuracy.Recently,CoupledMapping(CM)hasbeenintroducedintofacerecognitionframeworkacrossdifferentresolutions,whichlearnsacommonfeaturesubspaceforbothhigh-resolution(HR)andlow-resolution(LR)faceimage

5、s.Inthispaper,inspiredbymaximummarginprojection,weproposeLargeMarginCoupledMapping(LMCM)algorithm,whichlearnsprojectionstomaximizethemarginbetweendistanceofbetween-classsubjectsanddistanceofwithin-classonesinthecommonspace.ExperimentsonpublicFERETandSCfacedatabasesdemonstratethatLMCMiseffecti

6、veforlow-resolutionfacerecognition.Keywords:CoupledMapping
Low-resolutionfacerecognition
LargeMarginCoupledMapping
FERET
SCface1IntroductionAgreatnumberofachievementshavebeenmadeintheareaofautomaticfacerecog-nitionduringlastdecades,especiallyunderwell-controlledcircumstances.However,theperfor

7、manceoffacerecognitionsysteminrealworldalwaysdegradesdramaticallywhenthequalityofinputfaceimagesbecomespoor,suchaslow-resolution.Thisisaspeci?cconcerninsurveillanceenvironmentwherethetargetisfarfromthesensor,resultinginlow-resolutionfaceimage