Unconstrained Japanese address recognition using a combination of spatial information and word knowledge

We describe a new handwritten address recognition method which can correct the errors occurring in line extraction, character segmentation, and character recognition as a possible means of avoiding the error accumulation which occurs during the recognition sequence in conventional methods. We formulate the address recognition method as a minimum cost search problem. We define the character recognition cost which estimates the reliability of the character recognition result, the arrangement cost which estimates the plausibility of the character string´s spatial arrangement, and the word knowledge cost which estimates the plausibility of the linguistic conditions. By using a combination of these costs, the proposed method can recognize an address which has not been extracted as a single line from input images by a conventional method. The efficiency of the proposed method is evaluated through an experiment using 600 Japanese mail images. An address recognition rate of 79.38% was obtained