Algorithm using deep learning for recognition of japanese historical characters in photo image of historical book

Abstract

In Japan, there are vast amount of classical books written in cursive Japanese that cannot be read by modern people. It is difficult to recognize each cursive Japanese separately, because they are written connected. Furthermore, there are many types of shape of characters. Therefore, an efficient method to convert them into modern characters automatically is required. Some methods recognizing a block of a few characters using deep learning have been studied so far. However, every page in a Japanese historical book is stored by a photo image; therefore, it is desirable to recognize all characters in a photo image at once. In this paper, we propose a method using deep learning to recognize cursive Japanese in a photo image without separating a block of characters manually. Furthermore, we evaluate the performance of the proposed algorithm using photo images of an actual book.