On Feature Selection in Multiclass Pattern Recognition

摘要

Several possible solutions are presented to the problem of feature selection in multiclass pattern recognition. The principal objective of research reported in this thesis is to develop a generalized mathematical formulation of feature selection techniques in multiclass pattern recognition. It is also emphasized to investigate the practical applications of the proposed feature selection techniques by experiments of real data simulations from agricultural remote sensing. The probability of system error (or misclassification) is used as the index of feature effectiveness. For multiclass pattern recognition, the over-all system error could be treated in terms of pair-wise misclassifications, and the separability measure between a pair of classes is considered as the principal measure of the feature effectiveness. Using minimax linear discriminant functions, a separability measure is derived for normal class distributions with unequal covariance matrices. An optimum seeking procedure is developed for feature selection in multiclass pattern recognition. A non-parametric feature selection technique is proposed. It is hoped that a finite number of classes is represented by some finite number of unknown probability structures and are distributed in a finite discrete measurement space. Two existing measurement-space transformation methods for feature selection are reviewed, and their applications are tested by computer simulation of data from agricultural remote sensing. (Author)