Japanese Cursive Character Recognition for Efficient Transcription Kazuya Ueki and Tomoka Kojima School of Information Science, Meisei University, Tokyo, Japan Keywords: Character Recognition, Japanese Cursive Character, Kuzushiji, Convolutional Neural Network. Abstract: We conducted detailed experiments of Japanese cursive character recognition to promote Japanese historical document transcription and digitization by using a publicly available kuzushiji dataset released by the Cen- ter for Open Data in the Humanities (CODH). Using deep learning, we analyzed the causes of recognition difficulties through a recognition experiment of over 1,500-class of kuzushiji characters. Furthermore, assu- ming actual transcription conditions, we introduced a method to automatically determine which characters should be held for judgment by identifying difficult-to-recognize characters or characters that were not used during training. As a result, we confirmed that a classification rate of more than 90% could be achieved by narrowing down the characters to be classified even when a recognition model with a classification rate of 73.10% was used. This function could improve transcribers’ ability to judge correctness from context in the post-process—namely, the previous and subsequent characters. 1 INTRODUCTION As Japanese characters have changed considerably over time, it has become difficult for non-experts to read classical Japanese literature. By transcri- bing historical documents, we are able to understand the events of past eras. Therefore, research on the transcription and digitization of historical documents has been conducted. However, a vast number of li- terary works have not yet been digitized. One of the considerable barriers to this digitization is that Japa- nese classical literature was written in a cursive style using kuzushiji characters that are very difficult to read compared with the contemporary style. The primary difficulty that arises in kuzushiji cha- racter recognition is rooted in the fact that most kuzushiji characters are very difficult to distinguish clearly from other characters because the writing style varied among different eras and authors. Therefore, in this study, assuming the actual transcription process, we considered a method that could automatically de- tect which characters were difficult to classify using deep learning and mark them as unknown characters without making the final decision. 2 RELATED WORK Currently, handwritten characters can be recognized with high accuracy by using convolutional neural ne- tworks (CNNs). This trend has been highly motivated by a CNN called LeNet with a convolution and po- oling structure that was proposed by (LeCun et al., 1989)(Lecun et al., 1998); this network successfully recognized handwritten digits. CNNs have also been widely used for recognizing kuzushiji characters in classical literature and achieved relatively high accu- racy (Hayasaka et al., 2016)(Ueda et al., 2018). These studies on kuzushiji recognition have only focused on classifications of less than 50 hiragana character cla- sses. However, classical documents are not limited to hiragana, as they also include katakana and kanji characters. To create electronic texts of classical li- terature or documents, it is necessary to recognize a wide range of characters and improve the accuracy with which they are transcribed. Based on this, we focused on recognizing not only hiragana characters but also katakana and kanji characters using the publi- cly available kuzushiji dataset (Clanuwat et al., 2018). When a text includes many character classes, we need to address data imbalance problems. Yang et al. conducted experiments of hand-written characters from documents recorded by the government-general of Taiwan from 1895 to 1945 (Yang et al., 2019). The data imbalance problem was solved by applying data augmentation when training deep learning models. Reducing the labor load on transcribers was also considered an important issue. Thus, a new type of OCR technology was proposed by (Yamamoto and Osawa, 2016) to reduce labor in high-load reprint work. It was deemed important to divide reprinting work among experts, non-experts, and an automated 402 Ueki, K. and Kojima, T. Japanese Cursive Character Recognition for Efficient Transcription. DOI: 10.5220/0008913204020406 In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 402-406 ISBN: 978-989-758-397-1; ISSN: 2184-4313 Copyright c  2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved