End-to-End Pre-Modern Japanese Character (Kuzushiji) Spotting with Deep Learning

• Item type: Symposium(1)
• 公開日: 2018-11-24
• タイトル: End-to-End Pre-Modern Japanese Character (Kuzushiji) Spotting with Deep Learning
• タイトル / 言語: en
• タイトル: End-to-End Pre-Modern Japanese Character (Kuzushiji) Spotting with Deep Learning
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: Machine learning\n
• キーワード / 主題Scheme: Other
• キーワード / 主題: Neural network\n
• キーワード / 主題Scheme: Other
• キーワード / 主題: U-Net\n
• キーワード / 主題Scheme: Other
• キーワード / 主題: Kuzushiji\n
• キーワード / 主題Scheme: Other
• キーワード / 主題: Pre-modern Japanese books\n
• キーワード / 主題Scheme: Other
• キーワード / 主題: Character spotting
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: ROIS-DS-CODH / NII
• 著者所属: MILA / Université de Montréal
• 著者所属: ROIS-DS-CODH / NII
• 著者所属(英): en
• 著者所属(英): ROIS-DS-CODH / NII
• 著者所属(英): en
• 著者所属(英): MILA / Université de Montréal
• 著者所属(英): en
• 著者所属(英): ROIS-DS-CODH / NII
• 著者名: Tarin, Clanuwat; Alex, Lamb; Asanobu, Kitamoto
• 著者名(英): Tarin, Clanuwat; Alex, Lamb; Asanobu, Kitamoto
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Kuzushiji has been used as a cursive writing style in Japan for over a thousand years. However, following the reform of Japanese language textbooks in the year 1900, Kuzushiji was no longer taught in schools. Thus at present, nearly all Japanese natives cannot read books written or published before 150 years ago. As a result, many important pieces of Japanese history are only accessible to a handful of specially trained scholars. This has motivated the use of machine learning to read Kuzushiji. To accomplish this, we propose a new way of using the U-Net architecture from Deep Learning to spot characters in manuscripts. For other researchers to use our system to search for characters or words in manuscripts written with Kuzushiji, only an image of the page of text needs to be given to our system as input. Other Kuzushiji word spotting method have limitations in characters or words written in different Jibos (字母) or root characters in Hentaigana (or variant Kana). Our proposed U-Net character spotting system not only gives results with state-of-the- art accuracy, our system also resolves the Jibo problem, allowing our system to search for input characters in the text. While we present results for character spotting, we discuss keyword search as an exciting area for future work.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Kuzushiji has been used as a cursive writing style in Japan for over a thousand years. However, following the reform of Japanese language textbooks in the year 1900, Kuzushiji was no longer taught in schools. Thus at present, nearly all Japanese natives cannot read books written or published before 150 years ago. As a result, many important pieces of Japanese history are only accessible to a handful of specially trained scholars. This has motivated the use of machine learning to read Kuzushiji. To accomplish this, we propose a new way of using the U-Net architecture from Deep Learning to spot characters in manuscripts. For other researchers to use our system to search for characters or words in manuscripts written with Kuzushiji, only an image of the page of text needs to be given to our system as input. Other Kuzushiji word spotting method have limitations in characters or words written in different Jibos (字母) or root characters in Hentaigana (or variant Kana). Our proposed U-Net character spotting system not only gives results with state-of-the- art accuracy, our system also resolves the Jibo problem, allowing our system to search for input characters in the text. While we present results for character spotting, we discuss keyword search as an exciting area for future work.
• 書誌情報: じんもんこん2018論文集; 巻 2018, p. 15-20, 発行日 2018-11-24
• 出版者 / 言語: ja
• 出版者: 情報処理学会