A fuzzy model for unsupervised character classification

This paper presents a fuzzy-logic approach to the efficient unsupervised character classification in order to increase the robustness, correctness, and speed of a follow-up optical character recognition system. The classification procedures are split into two stages. The first stage separates the characters into seven categories based on the word structure of a text line. The second stage, referring to pattern matching, is to classify all the characters in each category of stage one into a different set of prototypes. The existing methods of similarity measures and their problems are investigated, and a nonlinear weighted similarity function is proposed. A fuzzy model of unsupervised classification, which is more natural to represent the library of prototypes, is defined and the weighted fuzzy similarity measure is extended. Several propositions of the features of the fuzzy model are discussed. Finally, a preclassifier to speed up the classification is presented. The small set of prototypes can be recognized and postprocessed much easier and more efficient.