Optimization of Classifiers under Small Sample Conditions

摘要

Results form optimization of classifiers on the basis of the Karunen-Loeve Transform (KLT), known also by the name, principle of component analysis (PCA), are presented. The vectors of the spectral components of KLT and a geometric measure of the Minkowski dissimilarity measure are selected as the optimization parameters. The result of the optimization is minimization of the average number of prototypes of classes required for correct identification of a tested image with specific probability. The number of prototypes is determined by means of procedures of sequential analysis.