A Study Of Kernel Two-dimensional Principal Component Algorithm And Its Application

As an important aspect of pattern recognition, face recognition plays an increasingly important role in security detection and video surveillance. In recent years, with the development of artificial intelligence technology,we pay more and more attention to face recognition technology. Various theories and algorithms are constantly improving. Subspace-based algorithm for its convenience, fast and acclaimed, of interest to many researchers. Face recognition is a method of pattern recognition.The purpose of which is to extract useful identifying features of face images, and according to the extracted facial features to build mode so as to achieve classification. One of the most critical issue is feature extraction, feature extraction method determines recognition rate. In this paper, we put the idea of matrix factorization into feature extraction algorithm, reducing the number of dimensional of kernel matrix in nonlinear feature extraction algorithms. The main results obtained are as follows:(1) Introduce the linear feature extraction algorithms based on subspace,including the principal component analysis and discriminant analysis two methods,one for vector and another for matrix. For vector are principal component analysis(PCA) and linear discriminant analysis(LDA). For the two-dimensional matrix are two-dimensional principal component analysis(2DPCA) and two-dimensional discriminant analysis(2DLDA).(2) Study on nonlinear feature extraction algorithms based on subspace. Including the kernel principal component analysis(KPCA) and kernel two-dimensional principal component analysis(K2DPCA) algorithms. For the large-scale training database,in order to overcome kernel matrix is too large to compute in KPCA and K2 DPCA algorithm, We use pivoted Cholesky decomposition to obtain approximation of kernel matrix to solve this problem. The experiment results in the noised ORL face database show that the process of the selected principal components can overcome the influence of noise in a certain extent. The recognition rate of KPCA and K2 DPCA compared with obviously improved.And in a large Extended Yale B face database experiment result also indicate that this algorithm can effectively extract nonlinear features.(3) Research nonlinear kernel discriminant analysis(KLDA)and two-dimensional kernel discriminant analysis(K2DLDA)algorithms based on subspace. For large-scale kernelmatrix, we approximate it by the pivoted Cholesky decomposition algorithm. We combinate K2 DLDA and the pivoted Cholesky decomposition algorithms to propose a nonlinear feature extraction method. The large-scale experiment on the subset of the AR face database indicate that this method for classification result can enhance in a certain degree, and the pivoting process to some extent overcome the noise.