Matrix Reduction Based on Generalized PCA Method in Face Recognition

Abstract

Human face recognition technology is one of the hottest research in the field of pattern recognition at present. In this paper, the principle component analysis (PCA) and bidirectional principle component analysis (BDPCA) methods are proposed to recognize a grayscale face image, for which the size of the spatial distribution is 64 ? 64. At first, the main part of the face is extracted to form the eigen face pace with K-L transform, and in this process, the Singular Value Decomposition (SVD) is proposed to solve the eigen value and eigenvector because of the overlarge dimension of the covariance matrix of the training image. Then, the testing image is projected onto the eigen face space to get a group of projection coefficients named eigenvectors, which are compared with eigenvectors of all training face images in the Euclidean distance and establish proper threshold values for the face recognition. When the testing image is recognized, human and inhuman face image are distinguished by PCA, and if that the input image is a human face is right, BDPCA, by which the dimension of the row and column in the training and the testing image is reduced, is used to prove whether the input image is in the face database by minimun Euclidean distance between the projective eigen matrix of the testing image and the projective eigen matrix of the training image. The numberous experimental results indicate that PCA algorithm and BDPCA algorithm are effective and BDPCA face recognition rate is higher than PCA face recognition rate, especially in the case of small number of train samples.