Handwritten character recognition employing topology, cross correlation, and decision theory

Abstract

This paper describes a pattern recognition method in which a nonlinear comparison is made between the topological and shape properties of an unknown character and standards determined in a learning phase. Each character is expressed as a vector. The dimensions of the vectors are the points defining the character in the order they are encountered. The coordinates are the complex coordinates of the points in the 2-dimensional grid system being used. The unknown is compared with each standard of the same topological type by measuring the vector distance between them. From this measurement and associated conditional probability factors, the probability that the unknown is of the class of the standard is determined, utilizing the Bayes rule. The standard yielding the highest probability identifies the unknown. The necessary factors to utilize the Bayes rule were determined from a set of 26 samples of each of the characters 1,…, 9, 0. 260 additional samples were then tested of which 98.5 percent were recognized correctly, 0.0 percent were not recognized, and 1.5 percent were recognized incorrectly, thus indicating the validity of the system. It is also concluded that proper selection of features to be compared is much more important than highly precise statistical calculations on the features.