Japanese Character Segmentation for Historical Handwritten Official Documents Using Fully Convolutional Networks

摘要

This paper proposes a character segmentation method using a fully convolutional network (FCN) and a post-processing phase. The network is trained with five-channel images that indicate five kinds of zones within the bounding box for each character-the top half, bottom half, left half, right half, and center. The post-processing step reconstructs the bounding boxes for characters from the five-channel image of the FCN output. The proposed method possesses the following advantages: (1) It is possible to process input images including multiple text lines directly; in other words, a text line segmentation process is unnecessary. (2) It does not rely upon character recognition. (3) It is robust to variations in the sizes of characters and the gaps between characters and also to cursive characters or character overlap. In the experiment of character segmentation, the accuracy ratio was 95% for real images of historical handwritten official documents written in Japanese.