Research On The Face Recognition Based On Nonlinear Feature Extraction And Neural Network

Feature extraction is one of the elementary problems in the area of pattern recognition. The extraction of the effective image features is the key step for image recognition tasks. Kernel-based feature extraction methods are a very effective nonlinear feature extraction method which is proposed by Vapnik. In this paper, we focus our attention on kernel-based nonlinear feature extraction and develop some new algorithms. These algorithms are verified to be effective in the application of face recognition. It is well known that kernel method is widely used in pattern recognition. A two-stage kernel feature extraction methods for face recognition is developed in this paper. The algorithm includes two stages: firstly, the classical principal component analysis (C-PCA) is employed to condense the dimension of image vector. What follows, Kernel principal component analysis (KPCA) are applied to the reduced dimensional training samples. On this basis a more efficient method, called I-PCA+KPCA, are proposed. Different from the previous method where C-PCA is based on vectors, I-PCA is to exploits image matrices to directly construct the image total scatter matrix. Finally, the experimental results on ORL face databases indicate that the proposed method is more efficient than KPCA while retaining the same recognition. Based on tow-stage kernel feature extraction method, a face recognition framework is proposed. After the features of the images are extracted, an improved back propagation (BP) algorithm is introduced to train the neural network for recognition. The framework combines the optimization of the I-PCA+KPCA and the adaptability of the neural network to improve the recognition rate and the robustness of the algorithm to noises. Comparing with the traditional PCA and KPCA algorithm, the method proposed here can significantly have a better effect on extracting representation of the differences among different faces, which results in a high recognition and a high robustness of the algorithm. Experimental results presented in this thesis verified that the proposed algorithm is accurate and effective.