A Neural Approach to Concurrent Character Segmentation and Recognition

This paper presents a neural network solution that combines character segmentation and character recognition concurrently as a single task. Current segmentation methods utilize traditional image processing techniques such as spatial histograms which are only 60% accurate on handprint. Using traditional techniques for segmenting handprint in a model recognition system running on a massively parallel machine requires 55% of the entire processing time while the neural network classification requires 0.34% of the time. A neural network based solution for segmentation offers improvements in both speed and accuracy. In order to demonstrate feasibility, initial experiments were conducted on machine printed digits. The results demonstrate that neural networks can be used for concurrent segmentation and recognition. Two different neural network solutions are studied, one based on a self-organized multi-map architecture, and the other based on the use of multi-layered perceptrons. Both approaches achieve 100% segmentation and recognition over a test set of 1,104 image samples. The multi-layered perceptron solution processes the activation signals from two separately trained networks whereas comparable results are achieved using the raw associations produced from a single self-organized network.