Recognition of Handwritten Chinese Postal Addresses Using A Dual-Expert Classification Scheme

Abstract

A mail shorting system based on the development of OCR system is proposed in the paper for the automatic mail processing. The system to recognize the handwritten Chinese postal addresses on standard envelopes is constructed by using dual-expert classification scheme. Each architecture of dual-expert classification scheme includes feature extraction stage, recognition stage, and semantic processing stage. The black local density feature (BLDF) and multi-layer percetron (MLP)are used in the first expert classification for handwritten Chinese characters (HCCs). The white local density feature (WLDF) and Bayesian network are used in the other expert classification for HCCs. Since the names of city on addresses contain contextual information, the performance of recognition can be increased by the semantic processing. Finally, the comparison of results from the two expert classifications is used to reduce error rate of address recognition. By using dual-expert classification, we could improve the performance of the system with 70.2% correction rate and 1.3% error rate in our experimental results.