Research Seminar - April 11, 2003

An Improved Pattern Classification System using Optimal Feature Selection,
Classifier Combination, and Subspace Mapping Techniques.

Abstract:

Pattern recognition is one of the key features of intelligent behavior for both humans and machines. The classification performance is an important aspect that affects the overall behavior of pattern recognition systems. The focus of this work is primarily on developing powerful feature selection, classifier combination and subspace mapping techniques. To solve the problem of feature selection, we discuss a new scheme based on the concept of mutual information in selecting the most important features for a given classification task. A new technique for combining classification results of different classifiers is also proposed. The technique is based on the Dempster-Shafer theory of evidence, which is a powerful method for representing uncertainties and lack of knowledge. In addition to the above, we propose a new subspace mapping method that is applicable to multi-channel signals and takes into consideration both intra- and inter-channel relationships. The method is based on maximizing the mutual information within and between different channels and is called the hybrid information maximization (HIM) algorithm. The proposed algorithms are implemented and tested on synthetic and real data. In particular, the algorithms are applied to speaker identification, multi-sensor EEG data classification, and texture image classification.