資源描述:
《基于稀疏保持判別嵌入的人臉識別.pdf》由會員上傳分享,免費在線閱讀,更多相關內(nèi)容在應用文檔-天天文庫。
1、第35卷第2期儀器儀表學報V01.35No.22014年2月ChineseJournalofScientificInstrumentFeb.2014基于稀疏保持判別嵌入的人臉識別水王國強,李龍星,郭曉波(1.洛陽理工學院計算機與信息工程系洛陽471023;2.大連理工大學機械工程學院CAD&網(wǎng)絡研究所大連116024;3.安陽工學院計算機科學與信息工程學院安陽455000)摘要:最近,人們對高維數(shù)據(jù)(例如人臉圖像)潛在的稀疏表征結(jié)構有很大興趣。提出一種稱為稀疏保持判別嵌入(SPDE)新降維算法,該算法在稀疏保持投影(SPP)的目標函數(shù)中增加
2、了改進的最大間距準則(MMMC)。SPDE保留了SPP的保持稀疏結(jié)構特性,利用了MMMC的全局判別結(jié)構。SPDE合并了稀疏準則和Fisher準則,具有更強的判別力,尤其訓練集小的時候,更適合于人臉識別任務。SPDE能夠自然地避免小樣本問題并且計算是有效的。在3個公共人臉數(shù)據(jù)庫(ORL、Yale以及FERET)上的實驗結(jié)果表明SPDE對人臉識別是有效的和可行的。關鍵詞:降維;稀疏保持投影;改進的最大間距準則;人臉識別中圖分類號:TP391.41文獻標識碼:A國家標準學科分類代碼:510.40Sparsitypreservingdiscrimi
3、nantembeddingforfacerecognitionWangGuoqiang,LiLongxing,GuoXiaobo(J.DepartmentofComputerandInformationEngineering,LuoyangInstituteofScienceandTechnology,Luoyang471003,China;2.InstituteofCAD&NetworkTechnology,SchoolofMechanicalEngineering,DalianUniversityofTechnology,Dalian1
4、16024,China;3.SchoolofComputerScienceandInformationEngineering.AnyangInstituteofTechnology,Anyang455000,China)Abstract:Recently,thereislotsofinterestinthepotentialsparserepresentationstructureofhigh-dimensionalitydata,suchasfaceimages.Inthispaper,anoveldimensionalityreduct
5、ionalgorithm,calledsparsitypreservingdiscriminantembedding(SPDE),isproposedinwhichthemodifiedmaximummargincriterion(MMMC)isaddedintotheobjectivefunctionofsparsitypreservingprojection(SPP).SPDEretainsthesparsitystructurepreservingcharacteristicofSPPandutilizestheglobaldiscr
6、iminativestructureobtainedfromMMMC.TheproposedSPDEalgorithmcombinesthespar-sitycriterionandFishercriterionandhasstrongerdiscriminatingpower;especiallywhenthesizeofthetrainingsetissmal1.mealgorithmismoresuitableforfacerecognitiontask.TheSPDEalgorithmcanavoidthesmallsamplesi
7、zeproblemnaturallyandthecomputationisefficient.Theexperimentresultsonthreepubliclyavailablefacedatabases(ORL.YaleandFERET)showthattheSPDEalgorithmiSeffectiveandfeasibleforfacerecognition.Keywords:dimensionalityreduction;sparsitypreservingprojection(SPP);modifiedmaximummarg
8、incriterion(MMMC);facerecognition問題廣泛使用的方法是降維,降維的目標就是把高維人1引言臉數(shù)據(jù)轉(zhuǎn)換成有意義的低維表征。主成分分析(principa