A New Method of Two-Dimensional Direct LDA and Its Application in Face Recognition
Dong Wang, Shunfang Wang · 2016
The two-dimensional direct Linear Discriminant Analysis (2D-DLDA) algorithm is based on the direct LDA and two-dimensional LDA. The algorithm retains the useful null space and uses the original two-dimensional image matrix directly while it does not pay much attention to the influence of the edge class and the overlap of the classes. So an improved method of 2D-DLDA is proposed in this paper. This new method redefines the between-class scatter matrix and uses the deformation of the Fisher criterion. Thus, the new method weakens the effect the edge classes have on the selection of the projection direction. Then the projection of the training samples should be calculated with some Fractional steps and the subspace would be redirected. As a result, it would avoid the serious overlap of the classes. The experiments based on the face recognition show that the new method is available.