Feature representation and discrimination based on Gaussian mixture model probability densities—Practices and algorithms

摘要

This study promotes the use of statistical methods in specific classification tasks since statistical methods have certain advantages which advocate their use in pattern recognition. One central problem in statistical methods is estimation of class conditional probability density functions based on examples in a training set. In this study maximum likelihood estimation methods for Gaussian mixture models are reviewed and discussed from a practical point of view. In addition, good practices for utilizing probability densities in feature classification and selection are discussed for Bayesian and, more importantly, for non-Bayesian tasks. As a result, the use of confidence information in the classification is proposed and a method for confidence estimation is presented. The propositions are tested experimentally.