Discriminant Analysis of Stochastic Models and Its Application to Face Recognition

As the vital component of a recently developed stochastic model based feature generation scheme, Fisher score is increasingly used in classification applications. In this work we present a generalization of previous proposed feature generation schemes by introducing the concept of multi-class mapping which is oriented to multi-class classification problems. Based on the generalized feature generation scheme, a novel face recognition system is developed by a systematical integration of hidden Markov model (HMM) and linear discriminant analysis (LDA). The proposed system is evaluated on a public available face database of 50 subjects. Comparing to holistic features based LDA method, stand alone HMM method, and LDA method based on previous proposed feature generation schemes which are intrinsically oriented to two-class problems, superior performance is obtained by our method in terms of recognition accuracy.