Discriminant feature analysis for pattern recognition

Abstract

Discriminant feature analysis is crucial in the design of a satisfactory pattern recognition system. Usually it is problem dependent and requires specialized knowledge of the specific problem itself. However, some of the principles of statistical analysis may still be used in the design of a feature extractor, and how to develop a general procedure for effective feature extraction always remains an interesting and also challenging problem.

In this thesis we have investigated the limitations of traditional feature extraction algorithms like Fisher's linear discriminant (FLD) and devised new methods that overcome the shortcomings of FLD. The new algorithm termed recursive cluster-based Bayesian linear discriminant (RCBLD) has a number of advantages: it has a Bayesian criterion function in the sense that the Bayes error is confined by a coherent pair of error bounds and the maximization of the criterion function is equivalent to minimization of one of the error bounds; it can deal with complex class distributions as unions of Gaussian distributions; it also has no feature number limitation and can fully extract all discriminant information available; the solution of the algorithm can be easily obtained without resorting to some gradient-based methods.

Since the proposed algorithms are designed as general-purpose feature extraction tools, they have been applied to a wide variety of pattern classification problems such as face recognition and brain-computer-interface (BCI) applications. The experimental results have verified the effectiveness of the proposed algorithms.