Face Recognition Analysis Based On Non-zero Subspace

Local linear embedding proposed that face-data would found in some low-dimensional subspace, all face-data would be linearity-denoted optimal with data in neighbourhood of the data.Though its drawback is mapping on training samples but not new test sample. Discriminant neighborhood embedding keep original local linear relationship of neighborhood in same classes and separate the samples in different classes by using the imformation of neighborhood and classes, so it has good recognition performance. Recognizing discriminant neighborhood embedding is obtained by linear projection vector of the optimal recognition, the vector carrying only the linear features. Problems in face recognition, as the external environment and face their own point of view, caused by the different expression of face images may not be linear separable.This linear discriminant analysis can not extract more discriminative nonlinear features. Fisher linear discriminant analysis of nuclear first through appropriate nonlinear separable nonlinear mapping of the original sample space to a linear separable transform high dimensional feature space, extracted by this method is more discriminative nonlinear discriminant features, and overcome the deficiencies of linear mapping theory. Many scholars tend to get optimal projection vector from zero subspace of kernel spread inner matrix, completely abandoned the non-zero spatial information, in fact, the divergence within the class space to carry non-zero matrix of the human face of information will help improve the recognition rate, or sometimes even more than zero space to carry important information. In this paper, the nonlinear separable nonlinear mapping the input space is mapped into high dimensional linear separable feature space, combined with local linear neighborhood embedding and discriminant subspace learning methods, proposed the kernel fisher non-linear discriminant analysis Non-zero space algorithm to reconstruct the kernel spread inner matrix, the matrix is almost a full rank. Get the optimal feature vector in Non-zero subspace of the matrix, in the yale face database to test the results of the null space and nuclear non-linear algorithms Fisher comparison test show that this method is effective.