Optical character recognition by a neural network

An optical character recognition (OCR) system, which uses a multilayer perceptron (MLP) neural network classifier, is described. The neural network classifier has the advantage of being fast (highly parallel), easily trainable, and capable of creating arbitrary partitions of the input feature space. Issues in design of the neural network that we examine include the selection of input features, the choice of network learning and momentum parameters, and the selection of training patterns. We also provide a detailed analysis of the learning parameters to provide insight into the MLP, and to suggest a mechanism to automatically tune these parameters. An OCR neural network classifier was trained to recognize characters from a large number of fonts, thereby approaching an omnifont environment. Samples were selected from over 200 fonts and 50 typical office documents, for a total of 110,000 training patterns. In order to evaluate the performance of the MLP classifier, a comparison is made with a high performance dynamic contour warping (DCW) classifier. The base recognition rate on the test set is 96.7% for the neural network classifier, compared to 95.9% for the DCW classifier.