Adaptive Feature Transformation for Classification with Sparse Representation

《Adaptive Feature Transformation for Classification with Sparse Representation》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、AdaptiveFeatureTransformationforClassificationwithSparseRepresentationYaxinSun,GuihuaWenTheauthorsarewiththeSchoolofComputerScienceandEngineering,SouthChinaUniversityofTechnology,Guangzhou510006,China. E-mail:sunyaxin2005@163.com,crghwen@scut.edu.cnAbstract—

2、Sparserepresentationbasedclassification(SRC)hasbeenshowntobeaneffectivemethodforfacerecognition.Furthermore,theinputfeaturesofmoreandmoreclassifiersareextractedbydimensionalreductionmethods.However,wefindthatthereconstructionabilityofabasisforatestingsamplei

3、srelatedwithcosinedistancebetweenthisbasisandthistestingsample,butmostdimensionalreductionmethodsarebasedonEuclideandistance.Obviously,agapisexistedbetweendimensionalreductionmethodsandSRC.Inthispaper,weproposeanadaptivefeaturetransformationbasedonself-tunin

4、gpointtopointdistances(SPPDAFT)totransformfeaturestoanewfeaturespace.TheSPPDAFTcanmakesthatthecosinedistancesamongsamplesinnewfeaturespaceincreasewiththeEuclideandistancesamongsamplesinoriginalspace.Asaresult,thereconstructionabilityofabasistoatestingsamplei

5、nnewfeaturespacewouldbeindirectlyrelatedwiththeEuclideandistancebetweenthisbasisandthissampleinoriginalspace,andthenthegapbetweendimensionalreductionmethodsandSRCcanbereduced.TheexperimentalresultsonbenchmarkdatabasesshowtheeffectivenessofSPPDAFT.IndexTerms—

6、SparseRepresentationClassification,DimensionalReduction,FeatureTransformation,FaceRecognitionI.IntroductionSparserepresentationhasrecentlybeenappliedtoavarietyofapplicationsincomputervisionandmachinelearning[1-5].Itssuccessisattributedtothefactthatthedimensi

7、onalityofsignalssuchasnaturalimagesisoftenmuchlowerthanthatwhichisobserved,andthusitoffersamorecompactyetbetterdescriptionofnaturalsignalsfortheaboveapplications[4].Toextendsparserepresentationtotheproblemsofclassification,Wright[36]proposedsparserepresentat

8、ionbasedclassification(SRC),whichobtainsgoodresultsonfacerecognition.However,therearethreedefectsforSRC.Firstly,duetotheovercompletecodebookandtheindependentcodingprocess,thelocalityandthesimila