Multicategory Classification of Patterns Represented by High-Order Vectors of Multilevel Measurements

A method is introduced for deciding the class of a pattern described by a vector of measured features. Memory requirements and the time required for classifying a pattern are not excessive, even when there are many classes and many multilevel feature measurements. Two initial steps use univariate distributions of the measurements to select, for a pattern in question, a small subset of neighboring classes. If the pattern is one used in deriving the distributions, its true class is certain to be in the selected subset. The final step distinguishes among the members, if there are more than one, of the subset by using discriminants derivable by classic methods. The technique is tested experimentally on a set of about 26 000 alphabetic characters of nine type fonts. The characters range widely in quality. One third are taken as training patterns from which histograms for the first two steps and discriminants for the third step are derived. The rest are used to test the method. The features used are the characteristic loci.