This paper proposes a method of document understanding system for handwritten address image. In a previous work, the performance of rough classification is low, and final outputs are not so accurate. The causes of these problems are low rate of character recognition and fault of knowledge-based processing. To avoid these problems, the proposed method uses multiple candidates for the character recognition and reselects most available one by using address phrases. Furthermore, by changing the priority in rough classification and giving the dynamical threshold in rough classification, the method can overcome the fault of the knowledge-based processing. In experimental results, the effectiveness of our method has been proven.
document understanding system
address reading
character segmentation