图嵌入投影非负矩阵分解图像特征提取方法.pdfVIP

图嵌入投影非负矩阵分解图像特征提取方法 王 娟 杜海顺 侯彦东 金 勇 河南大学图像处理与模式识别研究所 开封 475004 摘要 针对投影非负矩阵分解(PNMF)不能揭示数据流形几何结构和判别信息的问题，提出了一种图嵌入投影非 负矩阵分解(GEPNMF)特征提取方法。首先分别构建了描述数据流形几何结构和类间分离度的近邻图，然后采用它 们的拉普拉斯矩阵设计了一个图嵌入正则项，并将其与PNMF的目标函数融合，以构造GEPNMF的目标函数。在 GEPNMF目标函数中引入了图嵌入正则项，使求得的子空间能够在保持数据流形几何结构的同时，类间间距也最大。 另外，还在 目标函数中引入了一个正交正则项，以确保GEPNMF子空间基向量具有数据局部表示能力。对求解 GEPNMF目标函数的累乘更新规则(MUR)进行了详细的推导。在Yale和CMUPIE人脸数据库上的实验结果表 明，提出的图嵌入投影非负矩阵分解特征提取方法比PNMF更适用于解决分类问题。 人脸识别;特征提取;图嵌入;非负矩阵分解 TP391.4 A GraphEmbeddingProjectiveNon-negativeMatrixFactorizationMethodforImageFeatureExtraction WANGJuan DUHai-shun HOUYan-dong JINYong InstituteofImageProcessingandPatternRecognition,HenanUniversity,Kaifeng475004,China Abstract Toovercomethedisadvantagethatprojectivenon-negativematrixfactorization(PNMF)failstodiscoverthe intrinsicgeometricalanddiscriminatingstructure,anovelgraphembeddingprojectivenon-negativematrixfactorization (GEPNMF)wasproposedforimagefeatureextraction.Thepaperconstructedtwoadjacentgraphsthatareseparately usedtocharacterizetheintrinsicgeometricalstructureofdataandinterclassseparability.UsingtheLaplacianmatrices oftheadjacentgraphs,thepaperdesignedagraphembeddingregularizationthatincorporateswithPNMFsobjective functiontoconstructtheGEPNMFsobjectivefunction.Sincethegraphembeddingregularizationisadoptedbytheob jectivefunction,thelearnedsubspaceofGEPNMFcanpreservethedatageometricalstructurewhileitmaximizesthe marginsbetweendifferentclasses.Thatistosay,ithasmorediscriminability.Inaddition,thepaperintroducedanor thogonalregularizationintotheobjectivefunctiontoensurethelearnedbasestobeparts-based.Thepaperdeduceda multiplicativeupdaterule(MUR)tooptimizetheobjectivefunction.TheexperimentalresultsonYaleandCMUPIE faceimagedatasetssuggesttheeffectivenessofGEPNMF. Facerecognition;Featureextraction;Graphembedding;Non-negativematrixfactorizati