An Efficient LDA Algorithm for Face Recognition

It has been demonstrated that the Linear Discriminant Analysis (LDA) approach outperforms the Principal Component Analysis (PCA) approach in face recognition tasks. Due to the high dimensionality of a image space, many LDA based approaches, however, first use the PCA to project an image into a lower dimensional space or so-called face space, and then perform the LDA to maximize the discriminatory power. In this paper, we propose a new, unified LDA/PCA algorithm for face recognition. The new algorithm maximizes the LDA criterion directly without a separate PCA step. This eliminates the possibility of losing discriminative information due to a separate PCA step. We discuss the connection between the new algorithm and the traditionalPCA+LDA approach. We also prove that the new algorithm is equivalent to the eigenface (PCA) approach in a special case, where each person has only one sample in the training set. The feasibility of the new algorithm has been demonstrated by experimental results.