Face Recognition Based On Radial Basis Function Neural Networks Classifier

With the development of computer science and communications technology, people pay much more attentions to the secure authentication than ever. Therefore, biometrics, which is a new identification approach, becomes more and more important. Face recognition is one kind of biometrics technology, which is in accord with the act or practice of one human being that thinks the most. Thus, face recognition has always been a hotspot in this field. Whereas, applied in real world, face recognition technology faces various difficulties of which the high-dimension problem and the small sample size problem are the most notable two. After giving a detailed development history and main characteristics of face recognition, this thesis attempts to solve the small sample size problem based on nonlinear Fisher discriminant analysis and radial basis function neural networks classifier.The first issue of this thesis is the application of nonlinear technique– kernel trick– in face recognition. Fisher discriminant analysis is a classical statistical method both for dimension reduction and classification. But it cannot gain good results in nonlinear issues. Under the circumstances, the thesis applies the kernel method in face recognition. A kernel transform is performed. Then a new algorithm for feature extraction called kernel discriminative common vectors (KDCV) approach is described. KDCV extracts the common vector of each given class as optimal discriminative vector in the kernel-transformed space. It reduces the high dimension of the training sample set into (M-1)-dimension. The experimental results show that KDCV gives excellent performances both in recognition rates and in computation burden of the calculations.The second issue is a radial basis function (RBF) neural networks classifier. The RBF is a feed-forward artificial neural network as while as a general and efficient method as a classifier in pattern recognition problem. Its simple structure, fast training process and favorable generalization ability, make RBF neural network perform well in lots of fields especially in pattern classification and function approximation. The thesis focuses on how to select the hidden units, i.e. centers of the RBF Neural Networks classifier. Then a RBF classifier...