Face Recognition Using Newly Regularized LDA
Lee Hui Kueh, John Tark Lee, Kwon Soon Lee · 한국정보기술학회논문지 · 2010
A newly proposed FR (Face Recognition) approach with a weighted regularization parameter based on the conventional R-LDA (Regularized Linear Discriminant Analysis) method was presented in this paper. In the case of SSS (Small Sample Size), since the face feature space has typically a large number of pixels and the total number of training samples is less than the dimension of face feature space, all the scatter matrices of LDA are singular and its recognition accuracy is directly deteriorated. Therefore, it is impossible to apply the origina1 LDA algorithm to the FR. In this paper, it was attempted to optimize the revised Fisher's criterion with a weighted regularization parameter as a solution of the SSS problem. ORL (Olivetti Research Lab) database by using MATLAB was analyzed and the performance of face recognition was evaluated. The simulations were given by identifying the test samples as the some of the training subjects. In addition, the recognition performance of the proposed approach was compared to the ones of the well-known conventional methods such as Eigenfaces and R-LDA, which were established in this paper. The main interest idea of this paper was to demonstrate the superiority in robustness against the SSS problem of the proposed approach to the conventional Eigenfaces and R-LDA methods. The recognition rate of the proposed approach were fairly higher than others. In the future, the applicability of this proposed algorithm to the FR of the large sample database shall be discussed.