Automatic Font Generation For Early-Modern Japanese Printed Books

摘要

The National Diet Library makes images of early-modern Japanese printed books public on the Internet so that anyone can read the valuable books, which have been published during Meiji era to early Showa era. For the efficient and convenient use of the data, it is necessary to convert image data into text data. However, it is difficult to recognize character images of early-modern printed books correctly by existing OCRs. The multi-fonts character recognition method for early-modern printed books has been proposed. The method requires huge number of character images of those old books while the current methods to collect character images are very limited. In this paper, we propose a new method to complement learning data for the multi-fonts character recognition method by automatically generating character images with deep learning. For that purpose, we construct the neural network that automatically generates character images which have the features similar to the corresponding characters to be found in other early-modern Japanese printed books.