Deep Learning Based Ancient Asian Character Recognition
Masahiko ATSUMI, Syunsuke Kawano, Tomoki Morioka, Lin Meng · 2020
Ancient documents record a lot of information such as politics, economy, culture and so on, which may help people know the history and culture. However, a lot of ancient documents have not be organized yet. One of the reasons is some kinds of character are not be used now and only few of specialists can understand them, especially in Asian. These ancient Asian characters include Oracle Bone Inscriptions, Kuzushi characters, etc. Furthermore, these ancient documents are found one after another, letting document organization becomes a hard work, and aging processing increases the difficult of organization. This paper aims to recognize some kinds ancient character for helping people organize the ancient Asian documents by deep learning method. For achieving better accuracy, we applied the most state-of-the-art deep learning model -self-attention, which is different with current convolutional neural network. The experimental results shows the self-attention also achieves good accuracy in Oracle Bone Inscriptions and Kuzushi characters recognition, which also prove the effectiveness self-attention in ancient character recognition.