Feature Extraction based on Sub-Pattern Multi-Directional 2DLDA

A novel feature extraction method based on sub-pattern Multi-directional two-dimensional linear discriminate analysis (Sp-MD2DLDA) for face recognition is presented in this paper. In the proposed method, firstly, we apply directional 2DLDA (D2DLDA) to extract features in some initial directions, and then choose the effective directions from the initial directions for feature fusion after an evaluation. Secondly, divide the original images into small regions and apply D2DLDA to a set of partitioned sub-patterns to obtain features in the selected effective directions which complement each other. Finally, fuse these complementary features and use nearest neighbor classifier for classification. Since the proposed method not only can extract local features and reduce the impact of the variations in expression and illumination by dividing the original images into smaller sub-images, but also extract features in many more directions, we expect that it can improve the recognition performance. The experimental results on Yale and ORL databases show that the proposed SpMD2DLDA method has better classification performance than that of the other related methods.