Improved Two - Dimensional Direct Linear Discriminant Analysis Method And Its Application In Face Recognition

Two-dimensional linear discriminant analysis (2D-LDA) is an important method in face recognition field. Compared with the linear discriminant analysis (LDA) algorithm, the2D-LDA method uses the original image matrices to do the calculation. So the method avoids the curse of dimensionality problem to some extent and reduces the storage space and computed strength as well. Nowadays, some extension algorithms based on the2D-LDA have been proposed. And the two-dimensional direct LDA (2D-DLDA) is one of them. This paper conducted a research on the2D-DLDA algorithm and provided a series of enhanced algorithm which is based on the two-dimensional direct LDA. According to the characteristic of2D-DLDA, the new algorithm is based on the classical LDA algorithm and part of its improvement ideas. As a consequence, the new method can get the more discriminative information and achieve a higher recognition rate while the algorithm is applied to face recognition. The main work can be described as follows:1. The paper improved the2D-DLDA algorithm and provided the2D-VFDLDA algorithm. By redefining the between-class scatter matrix and using the expansion form of the Fisher criterion, the new method can weaken the effect the edge classes have on the selection of the optimal projection direction. Besides, the method adopts the multistep Fractional operation instead of the discard operation directly in the projection of the samples. And in this way, the overlapping classes in the projection space can separate from each other. As a result, the new algorithm can perform better in the face recognition. In addition, the experiments based on the variant algorithm of the two-dimensional direct LDA in the ORL database show the effectiveness of the new method.2. Because of the existence the abnormal points in the image picture, the paper introduces the fuzzy thought to the2D-VFDLDA algorithm. And the new improved algorithm uses the fuzzy membership function to describe the two-dimensional image. So the method could get a more accurate class center and weaken the negative effect the abnormal points have on the recognition rate of the face recognition. Moreover, some experiments based on the other fuzzy linear discriminant analysis are performed under the same condition. The results are compared with that of the fuzzy2D-VFDLDA algorithm. And the results show that the improved algorithm is still superior to the other fuzzy algorithms in some degree.