关键词: 人脸识别, 特征选取, 子空间学习, L_2,1范数, 核判别分析

Abstract: Feature selection and subspace learning are two key problems in face recognition. To select the rich nonlinear features more accurately in face image and solve the small sample size problem, this paper proposes a new generalized kernel discriminant analysis based on L_2,1-norm regularization (L21GKDA). The proposed method implicitly maps the original samples into feature space by using kernel function, and obtains the generalized kernel Fisher criterion. Then it presents an efficient transformation, transforming its nonlinear model into linear regression model. In order to perform feature selection and subspace learning simultaneously, an L_2,1-norm penalty term is added to the objective function, and the  solution algorithm of the regularization method is also obtained. Due to the feature selection capability of L_2,1-norm penalty term, the recognition performance is greatly improved. Experiments on ORL, AR, PIE standard face databases illustrate that the new method can effectively select the nonlinear features of the face data, and improve the discriminant ability.

Key words: face recognition, feature selection, subspace learning; L_2,1-norm, kernel discriminant analysis