Writer independent and writer adaptive neural network for on-line character recognition.

I. Guyon, D. Henderson, P. Albrecht, Y. Le Cun, and J. Denker.
In S. Impedovo, editor, From pixels to features III , pages 493--506, Amsterdam, Elsevier.

A system for on-line recognition of characters entered on a touch-sensitive pad was designed. It is based on a Time Delay Neural Network (TDNN) which processes the pen trajectory by extracting and combining local topological features. The TDNN was first trained on a large number of writers to recognize digits and uppercase letters. Classical back-propagation training was improved to reduce the error rate on atypical writing styles such as those of left handed people. After this writer-independent training, the TDNN was combined with a novel postprocessor which allows adaptation to peculiar writing styles and fast writer dependent learning of new letters and symbols.

[ next paper ]