A comparison of two neural networks for hand-printed character recognition

The paper explores the potential of neural networks in improving the state of the art in hand-printed character recognition (HPCR) using real-world data. Considerable attention is devoted to the pre-processing of the image, and the extraction of features, as this is at least as important as the final classifier. Careful attention is given to the interaction between these stages, particularly in the adaptive feedback net, where the output from the classifier is used to modify the characteristics of the feature detectors during recognition. This is a new architecture, and further details are given in Banks and Elliman (1989). The other network architecture used was a layered feedforward net, often called a multilayer perceptron, trained by error back-propagation as described by Rumelhart, Hinton and Williams (1986). They test the two approaches on the set of all digits and upper-case letters