A fast algorithm for the minimum distance classifier and its application to Kanji character recognition

Abstract

A fast algorithm for the minimum distance classifier (MDC) is proposed. The MDC has been used in various areas of pattern recognition because it is simple and fast compared with other complicated classifiers. The algorithm proposed is much faster than the exhaustive one that calculates all the distances straighforwardly. Our algorithm, which produces the same output as the exhaustive, omits redundant calculations according to Karhunen-Loeve expansion. From the KL-expansion of the prototype patterns, we form a subspace of the feature space, in which the order of examining the prototypes is decided adaptive to a given unknown pattern. We have applied the algorithm to recognition of handprinted Kanji characters and measured its performance on the ETL9B database. As a result, the theoretical and practical speedups were 10-20 and 4-9, respectively.