The effectiveness of features in pattern recognition
S. Ray · Spiral (Imperial College London) · 1985
Feature evaluation criteria are investigated in the context of pattern recognition.This investigation is composed of three broad parts: critical examination of the existing methods, proposition of new methods based on the Mahalanobis distance, and empirical study of existing and proposed methods.Various two-class and multiclass probabilistic criteria, most of which originate from the concepts of overlap and/or distance between classes, are examined for their comparative assessment as measures of feature effectiveness.Bayesian error probability being an optimum measure of the performance of a pattern recognition system, the different methods are judged depending on their relationship with this error probability.The 'two-class measures considered include the Bhattacharyya coefficient, the Matusita distance, the divergence function, the Kolmogorov variational distance, the generalized separability measure of Lissack and Fu and the Mahalanobis distance.The multiclass measures include Matusita's measure of affinity.Shannon's conditional entropy, the Bayesian distance of Devijver, the reordered by D -based criteria in stage II Confusion matrix obtained by using the first 15 2 , .features of the ordering D^(2) and adopting the 158 159 163 168 171 176 leave-one-out principle 179 17 -CHAPTER 1* Measurement vector may be considered to be the initial feature vector, thus generalizing the concept of a 'feature vector'.** If the features are Gaussianly distributed then most of the mathematical techniques also provide bounds to the probability of error.