SwiftPost: a vision-based fast postal envelope identification system
SMC'09 Proceedings of the 2009 IEEE international conference on Systems, Man and Cybernetics
Hi-index | 0.00 |
The main contribution of the paper is that it presents a suffix tree based data structure for automatic handwritten Chinese address reading. Since lots of papers have dis- cussed the destination address block (DAB) location for Chinese, we will not extend it in this paper. Instead, we pay more attention to improve the address matching per- formance after DAB location. As some conventional meth- ods, the extracted text lines are pre-segmented into a se- ries of radicals. We then build a hierarchical structure of sub-strings from the recognized characters of valid radical combinations. Coarse address candidates are selected at the same time. In address maching, we incorporate post- code information to filter redundant addresses. The pre- segmented radicals are compared with candidate address and a cost function combining recognition and structrual cost is evaluated for final decision. In the system, character segmentation, recognition, string searching and matching are considered synchronously by taking advantage of lex- icon knowledge. Suffix tree can greatly facilitate the sub- string generation process and enable the matching process to start from any character to collect potentially bitty in- formation. Therefore, our algorithms is more robust to the intervening noises and irregular writing styles. Finallly, we test 1,000 handwritten Chinese envelopes and achieve a correct rate of 85.30% in 3.0 seconds per mail averagely.