A highly accurate method based on phased search of candidate rectangle lattice for handwritten address recognition

摘要

An address consists of two elements: place strings and street numbers. In the study of handwritten address recognition, it is common to recognize place strings from its image. However, this issue has not been directory discussed in literature. Many recognition systems require effective post-processing in address recognition. In case of using candidate rectangle lattice and its character recognition results, the post-processing might require huge amount of processing time. Consequently, this fact may ultimately force the system to ignore some parts of the place strings, and hence affects the performance of the system. We remark that the reading accuracy falls down when we expand the domain of the targeted image. In this paper, we propose a method based on phased search of candidate rectangle lattice for recognizing place strings with high precision. Applying our proposal to the IPTP 1840 image data of address strings, experimental results show clearly the efficiency of our proposed system in handwritten address recognition.