A Recognition Method Using Neighbor Dependence

Abstract

Within the framework of an early paper1 which considers character recognition as a statistical decision problem, the detailed structure of a recognition system can be systematically derived from the functional form of probability distributions. A binary matrix representation of signal is used in this paper. A nearest-neighbor dependence method is obtained by going beyond the usual assumption of statistical independence. The recognition network consists of three levels-a layer of AND gates, a set of linear summing networks in parallel, and a maximum selection circuit. Formulas for weights or recognition parameters are also derived, as logarithms of ratios of conditional probabilities. These formulas lead to a straightforward procedure of estimating weights from sample characters, which are then used in subsequent recognition. Simulation of the recognition method is performed on a digital computer. The program consists of two main operations-estimation of parameters from sample characters, and recognition using these estimated values. The experimental results indicate that the effect of neighbor dependence upon recognition performance is significant. On the basis of a rather small sample of 50 sets of hand-printed alphanumeric characters, the recognition performance of the nearest-neighbor method compares favorably with other recognition schemes.