An implementation of postal numerals segmentation and recognition system for Chinese business letters

In this paper, a practical postal numeral segmentation and recognition system for Chinese business letters is presented. Line information for the address blocks is gained from the envelope image by projection, then the sub-image of the postal numerals is segmented. The numerals are recognized by three independent classifiers, and their results are combined to form final results. An improved voting principle based on a priori knowledge of each classifier is utilized to combine multiple classifiers in order to achieve perfect performance for high recognition rate and low error rate. The recognition ability of each classifier, acting with the voting principle, is obtained by statistics for a large set of samples. Different thresholds gained by training are employed for different classes, which can yield better results. A parallel structure hardware is designed to implement a real-time image processing system. The performance of the system is shown to be satisfactory, sufficient for the requirements of practical use.