Learning a Sequence of Cursive-Style Japanese Characters in Classical Literary Works

This paper presents a method for analyzing and recognizing Kuzushiji characters in images using Transformer models. Kuzushiji are cursive-style Japanese characters historically employed in the writing and printing of Japanese classical literature from the eighth to the nineteenth century. Japanese characters are categorized into Hiragana, Katakana, and Kanji characters. Hiragana and Katakana are phonetic symbols, each comprising approximately 50 classes of basic symbols in contemporary Japanese. Classical Kuzushiji writings, however, include more than just the basic forms of Hiragana and Katakana, as some of them originated from modifications of different Chinese characters with similar phonetic sounds. Additionally, Kuzushiji writings incorporate numerous Kanji characters, which number in the thousands and are individually simplified and altered. In addition to such numerous classes of characters in Kuzushiji, these characters are written in a cursive style, where strokes within a single character and between characters are interconnected. Our approach for recognizing cursive-style Kuzushiji characters employs Transformer architecture, as it is well-suited for analyzing characters that are vertically connected without explicit borders or spaces. We anticipate that the attention mechanism within the Transformer will effectively learn the relationships among adjacent characters, thereby aiding in the accurate separation and identification of character regions and boundaries. Our project has just started, but we have already achieved 90% accuracy for transcription of pre-modern literary works in our experiments.