Design of a neural network character recognizer for a touch terminal.

I. Guyon, P. Albrecht, Y. Le Cun , J. Denker, and W. Hubbard.
Pattern Recognition, 24(2):105--119.
1991



We describe a system for on-line recognition of hand-printed digits and uppercase letters. Characters are entered on a touch terminal consisting of a transparent touch-sensitive screen overlayed on a liquid-crystal display. Drawing actions are recorded as a sequence of coordinates [x(t), y(t)]. Trajectories, representing single characters, are resampled with a fixed number of regularly spaced points. Coordinates are normalized to obtain invariance with respect to position and scale. Further preprocessing extracts local geometric information such as the direction of the movement, and the curvature of the trajectory. The final output of the preprocessor is a sequence of 81 vectors with 7 elements each. This sequence is then processed by a novel Time Delay Neural Network (TDNN). The TDNN is a multi-layer feed-forward network, the layers of which perform successively higher-level feature extraction and the final classification. TDNN's, which were previously applied to speech recognition, are well suited to sequential signal processing. This allows us to use a representation which preserves the sequential nature of the data, in contrast with other approaches based on pixel-map representation.

The same network was trained to recognize either digits or capital letters with a modified version of the back-propagation algorithm. The training set contained 12,000 examples produced by a large number of different writers. The error rate was 3.4% on 2,500 test examples from a disjoint set of writers. When allowed to reject 7.2%, the system made 0.7% substitution errors. The recognizer was implemented on a AT&T 6386 PC with an auxiliary AT&T touch terminal. The throughput of the system, including acquisition, preprocessing and display, was 1.5 character per second.

Keywords: character recognition, on-line character recognition, handwritten characters, neural networks, touch terminal, touch screen.



[ next paper ]