Calculation of multi-category minimum distance classifier recognition error for binomial measurement distributions

摘要

This paper presents an algorithm for calculating recognition error for minimum Hamming distance classifiers, a special case of the Bayes (optimum) classifier under certain constraints. The error rate algorithm is derived for the two-category case when the binary components of the measurement vector are binomially distributed. The algorithm is easily extended to the multi-category case when the ratio of total measurements to measurements used per dichotomization is large.A number of categorizers were designed using conventional methods and actual quantized typewritten characters. The recognition error was calculated: 1.(1) theoretically, using the algorithm; and2.(2) experimentally, using an independent test set of characters for the categorizers. Both sets of results are presented for comparison, verifying the ability of the algorithnm to predict recognition error of categorizers.