Comparison of two different quadratic pattern classifiers using the same test data

The paper compares two types of quadratic pattern classifiers which are derived from different points of view. The first one (concept A) is the Bayes classifier, optimum for multivariate normal distributions of pattern classes, proceeded by a truncated orthogonal transformation. The second one (concept ΒJ is the polynomial classifier restricted to the second degree and adapted in the mean square sense. Both concepts are worked out and adapted to the common task by two different institutions. The comparison is based on a rather voluminous sample of handprinted numerals, collected from more than 200 people. This sample set is intended to be used by other institutions for further comparisons. The important result of the experiments carried out is that both concepts require approximately the same number of coefficients for identical classifier performance. With respect to the computing effort necessary for one single classifier decision, however, concept Β is superior to concept A, whereby concept A in contrast to concept Β has not yet been optimized with regard to computing effort. © 1974 De Gruyter Oldenbourg. All rights reserved.