Character Recognition Using A Modular Spatiotemporal Connectionist Model

We describe a connectionist model for recognizing handprinted characters. Instead of treating the input as a static signal, the image is scanned over time and converted into a time-varying signal. The temporalized image is processed by a spatiotemporal connectionist network suitable for dealing with time-varying signals. The resulting system offers several attractive features, including shift-invariance and inherent retention of local spatial relationships along the temporalized axis, a reduction in the number of free parameters, and the ability to process images of arbitrary length. Connectionist networks were chosen as they offer learnability, rapid recognition, and attractive commercial possibilities. A modular and structured approach was taken in order to simplify network construction, optimization and analysis. Results on the task of handprinted digit recognition are among the best report to date on a set of real-world ZIP code digit images, provided by the United States Postal Service. The system achieved a 99.1% recognition rate on the training set and a 96.0% recognition rate on the test set with no rejections. A 99.0% recognition rate on the test set was achieved when 14.6% of the images were rejected.