非負矩陣分解方法及其在人臉識別中的應(yīng)用.pdf

1、蘭州理工大學(xué)碩士學(xué)位論文非負矩陣分解方法及其在人臉識別中的應(yīng)用姓名：郭建虎申請學(xué)位級別：碩士專業(yè)：計算機應(yīng)用技術(shù)指導(dǎo)教師：張永201004201F負矩陣分解方法及其在人臉識別巾的應(yīng)用曼曼曼璺皇曼曼曼曼皇曼曼曼曼曼皇曼曼曼曼曼曼maalal__IIIIII。AbstractNonnegativeMatrixFactorization(NMF)isarecentpowerfulmatrixdecompositiontechniquethat

2、approximatesanonnegativeinputmatrixbyalowrankapproximationcomposedofnon—negativefactorsIthasfoundwideapplicabilityacrossabroadspectrumfields，rangingfromproblemsinfacerecognition，textanalysis，andgenemicroarrayanalysis，tom

3、usictranscriptionHowever，MostNMFalgorithmssufferfromlackofconvergence，andwhentheyconverge，arenotoriouslyslowtoconvergeInparticular，whenappliedtofacerecognition，NMFgivespoorrecognitionperformanceInshort，themajorcontributi

4、onsofthisthesisaresummarizedasfollow：1ThisthesisexploresANLSPG—NMF(AlternatingNonnegativeLeastSquaresusingProjectedGradientwithArmijuleforNMF)algorithmWhilethepopularLeeandSeung’Smultiplicativeupdatemethodstilllacksconve

5、rgenceresults，ANLSPGNMFalgorithmexhibitsniceoptimizationpropertiesHowever，ANLSPGNMFalgorithmconvergesslowerthantheLeeandSeung’SmultiplicativeupdatemethodWhenusedforcomputingNMFANLSPGNMFalgorithmcallsprojectedgradientwith

6、Armijoruleperiterationforsolvingtwonon。negativeconstrainedlinearleastsquaresproblemsSelectingthestepsize僅isthemosttimeconsumingoperationintheprojectedgradientwithArmijuleAnewstrategyforselectingthestepsizeQisintroducedAm

7、odifiedANLSPG—NMFalgorithmisproposedTheexperimentalresultshowsthatthemodifiedANLSPGNMFalgorithmconvergesfasterthantheANLSPGNMFalgorithmThisnewmethodisthusallattractiveapproachtocomputingNMF2Thisthesisalsoexploresaweighte

8、dmodificationofFisherNonnegativeMatrixFactorization(FNMF)algorithmforfacerecognitionInspiredbythefactthatEyes，mouthandnosehavebeendeterminedtobeimportantforfaceperceptionandrecognitionandthesefeaturesareapproximatelyloca

9、tedinthecenterofaface，afeatureextractionmethodshouldgivemoreimportancetothecentralareaofeachfaceoftrainingimagesetAweightedmodificationofFNMFforimprovingFNMFbasedfacerecognitonperformanceisproposedbyintroducingaweightmat

10、rixatthecostfunctionoftheFNMFdecompositionTheexperimentalresultalsosupportstheconclusionthatthenewalgorithmcanachieveabetterperformanceinfacerecognitionKeywords：Non—negativeMatrixFactorization；FaceRecognition；Projectedgr