Algorithm For Feature Extraction In Face Recognition

Main content of this thesis includes face detection in color space and face recognition based on multivariate statistics analysis. In the thesis, we try to integrate the technique of multivariable statistical, neural network and statistical physics and the theory of intelligent image processing in computer vision, and puts forward and realizes a set of method.The core of face recognition is the selection of face representation and match strategy. Because of the robust of the algebra feature vector, the thesis mainly presents several this kind methods of face recognition.The most of the research on ICA is concentrated on the noise-free ICA model definition. The estimation of the noiseless ICA model seems to be a challenging task in itself, and thus the noise is usually neglected in order to obtain tractable and simple results. Moreover, it may be unrealistic in many cases to assume that the data could not be noise. The parameter in the noisy ICA model is more difficult estimation than that of the noise-free ICA model, and the thesis presents a method based on mean field approximation to get the noisy ICA model parameters. Mean field approximation, which is originated in statistical physics, has been frequently used in practical situations in order to evaluate expectations of state variables from the model parameters. There has not restriction in noisy ICA model that the number of observed data must be at least as large as the number of independent components. The experiments by speech and sound have validated the method can solve the over-complete case. Moreover, if we restrict the source and Mixture matrix is non-negative, the independent components can represent the independently well. The recognition results will further improve.