A Semi-Supervised Learning Framework for Handwritten Text Recognition using Mixed Augmentations and Scheduled Pseudo-Label Loss

• Item type: SIG Technical Reports(1)
• 公開日: 2023-02-23
• タイトル: A Semi-Supervised Learning Framework for Handwritten Text Recognition using Mixed Augmentations and Scheduled Pseudo-Label Loss
• タイトル / 言語: en
• タイトル: A Semi-Supervised Learning Framework for Handwritten Text Recognition using Mixed Augmentations and Scheduled Pseudo-Label Loss
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Tokyo University of Agriculture and Technology
• 著者所属: Tokyo University of Agriculture and Technology
• 著者所属: Tokyo University of Agriculture and Technology
• 著者所属: Hitachi Ltd.
• 著者所属: Hitachi Ltd.
• 著者所属: Hitachi Ltd.
• 著者所属: Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Hitachi Ltd.
• 著者所属(英): en
• 著者所属(英): Hitachi Ltd.
• 著者所属(英): en
• 著者所属(英): Hitachi Ltd.
• 著者所属(英): en
• 著者所属(英): Tokyo University of Agriculture and Technology
• 著者名: Masayuki, Honda; Hung, Tuan Nguyen; Cuong, Tuan Nguyen; Cong, Kha Nguyen; Ryosuke, Odate; Takashi, Kanemaru; Masaki, Nakagawa
• 著者名(英): Masayuki, Honda; Hung, Tuan Nguyen; Cuong, Tuan Nguyen; Cong, Kha Nguyen; Ryosuke, Odate; Takashi, Kanemaru; Masaki, Nakagawa
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: We propose Incremental Teacher Model, a semi-supervised learning (SSL) framework for handwriting text recognition. The framework comprises a teacher model and a student model learning through Scheduled Pseudo-Label loss with Mixed Augmentations. First, the student model is pre-trained by labeled samples and used to initiate the teacher model. The student model is further trained on transformed samples provided by Mixed Augmentations. This training process uses pseudo-labels generated by the teacher model. After a training epoch, the teacher model is updated from a well-validated student model. We apply the proposed framework to four architectures of different handwriting recognizers. For almost every architecture, the recognizer trained by Incremental Teacher Model outperforms the recognizers trained by the well-known SSL methods on the IAM handwriting database.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: We propose Incremental Teacher Model, a semi-supervised learning (SSL) framework for handwriting text recognition. The framework comprises a teacher model and a student model learning through Scheduled Pseudo-Label loss with Mixed Augmentations. First, the student model is pre-trained by labeled samples and used to initiate the teacher model. The student model is further trained on transformed samples provided by Mixed Augmentations. This training process uses pseudo-labels generated by the teacher model. After a training epoch, the teacher model is updated from a well-validated student model. We apply the proposed framework to four architectures of different handwriting recognizers. For almost every architecture, the recognizer trained by Incremental Teacher Model outperforms the recognizers trained by the well-known SSL methods on the IAM handwriting database.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2023-CVIM-233, 号 44, p. 1-6, 発行日 2023-02-23
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8701
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会