Application of linear weight neural networks to recognition of hand print characters

In business and government institutions, a large amount of information is gathered from pre-printed forms completed in hand print. In many cases, this information must then be manually keyed into computers by human operators, causing a major bottleneck in the form processing chain. Automatic input of the information using optical character recognition (OCR) techniques is often impractical, due to the poor performance of such techniques on hand print. This is due to the high variability of the characters between examples from different people, and between examples from the same person taken at different times. In this paper, we discuss the use of a linear-weight neural networks based on Volterra functions for the recognition of hand-printed characters from pre-printed forms. We show that, by the use of a novel two-stage network architecture and high-speed DSP chips, levels of recognition performance and speed can be achieved that make automatic input of hand print to a computer reliable and practical