Self-supervised Contrastive Learning Using Triplet Loss for Offline Recognition of Handwritten Chinese Text lines

• Item type: SIG Technical Reports(1)
• 公開日: 2022-03-03
• タイトル: Self-supervised Contrastive Learning Using Triplet Loss for Offline Recognition of Handwritten Chinese Text lines
• タイトル / 言語: en
• タイトル: Self-supervised Contrastive Learning Using Triplet Loss for Offline Recognition of Handwritten Chinese Text lines
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション5-A
• 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: Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者所属: Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者所属: Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者所属(英): en
• 著者所属(英): Department of Computer and Information Science Tokyo University of Agriculture and Technology
• 著者名: Trung, Tan Ngo; Hung, Tuan Nguyen; Masaki, Nakagawa
• 著者名(英): Trung, Tan Ngo; Hung, Tuan Nguyen; Masaki, Nakagawa
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: In this paper, we propose a framework for contrastive learning of visual representations using online triplet loss and apply it for offline recognition of handwritten Chinese text lines. In this framework, the visual encoder model is trained with unlabeled text line images, then finetuned on ones with labels. As far as we know, it is the first approach that uses self-supervised contrastive learning for Chinese text line recognition. We apply the CRNN model to recognize text line images. At first, only the CNN part is trained in the proposed framework, and then it is used as the initial weight for the CRNN model when finetuned. In the experiments, we evaluated the performance of the proposed framework on the CASIA dataset. The results show that the text line recognizer trained with the self-supervised pre-trained encoder has outperformed the one without the pre-trained model.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: In this paper, we propose a framework for contrastive learning of visual representations using online triplet loss and apply it for offline recognition of handwritten Chinese text lines. In this framework, the visual encoder model is trained with unlabeled text line images, then finetuned on ones with labels. As far as we know, it is the first approach that uses self-supervised contrastive learning for Chinese text line recognition. We apply the CRNN model to recognize text line images. At first, only the CNN part is trained in the proposed framework, and then it is used as the initial weight for the CRNN model when finetuned. In the experiments, we evaluated the performance of the proposed framework on the CASIA dataset. The results show that the text line recognizer trained with the self-supervised pre-trained encoder has outperformed the one without the pre-trained model.
• 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 2022-CVIM-229, 号 31, p. 1-6, 発行日 2022-03-03
• 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