Two-dimensional direct discriminant locality preserving projection analysis for face recognition

To make use of information contained in the null space of withinclass during the implementation of discriminant locality preserving projection(DLPP), a very efficient feature extraction algorithm called two-dimensional direct discriminant LPP (2D-DDLPP) algorithm is proposed for face recognition in this paper. By modifying the simultaneous diagonalization procedure, the null space of the interclass matrix can be discarded for it carries no discriminative information and the null space of intraclass matrix is preserved for it contains very important information for classification. Also, the 2D-DDLPP algorithm does not need to transform 2D image matrix into a vector prior to feature extraction so that it can be implemented more efficient and accurate than the 1D traditional in extracting the facial features. Therefore, the performance of 2D-DDLPP has been greatly improved. Extensive experiments are performed to test and evaluate the new algorithm using the UMIST and the AR face databases. The experimental results indicate that our proposed 2D DDLPP method is not only computationally more efficiently but also more accurate than the 2DLPP method in extracting the facial features for face recognition.