Visual Constraints for Generating Multi-Domain Offline Handwritten Mathematical Expressions

• Item type: SIG Technical Reports(1)
• 公開日: 2022-03-03
• タイトル: Visual Constraints for Generating Multi-Domain Offline Handwritten Mathematical Expressions
• タイトル / 言語: en
• タイトル: Visual Constraints for Generating Multi-Domain Offline Handwritten Mathematical Expressions
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション3-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: The National Center for University Entrance Examinations
• 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
• 著者所属(英): en
• 著者所属(英): The National Center for University Entrance Examinations
• 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: Huy, Quang Ung; Hung, Tuan Nguyen; Cuong, Tuan Nguyen; Tsunenori, Ishioka; Masaki, Nakagawa
• 著者名(英): Huy, Quang Ung; Hung, Tuan Nguyen; Cuong, Tuan Nguyen; Tsunenori, Ishioka; Masaki, Nakagawa
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Offline Handwritten Mathematical Expression (HME) recognition has been intensively studied for two decades. However, most studies have worked on rendered images of online HMEs (Rendered Domain, or RD in short), but not on optically scanned HME images (Optical Domain, or OD in short). Due to a large gap between these two domains, it is challenging to use an HME recognizer trained in RD to recognize patterns in OD. To utilize a recognizer of RD in recognizing OD patterns, we propose a visual constrained CycleGAN (CCycleGAN) model to generate synthetic OD patterns from RD patterns and vice versa. For the RD-to-OD direction, we train a new recognizer of OD using the synthetic patterns. For the OD-to-RD direction, we utilize a pre-trained HME recognizer of RD to recognize the synthetic RD patterns generated from the OD patterns. Our experiments show that the CCycleGAN model performs significantly better than the CycleGAN model in terms of recognition accuracy. The recognition results demonstrate the effectiveness of our method in recognizing OD patterns using labeled RD patterns and unlabeled OD patterns for training. In addition, our CCycleGAN can well perform even when using only 100 training patterns in OD.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Offline Handwritten Mathematical Expression (HME) recognition has been intensively studied for two decades. However, most studies have worked on rendered images of online HMEs (Rendered Domain, or RD in short), but not on optically scanned HME images (Optical Domain, or OD in short). Due to a large gap between these two domains, it is challenging to use an HME recognizer trained in RD to recognize patterns in OD. To utilize a recognizer of RD in recognizing OD patterns, we propose a visual constrained CycleGAN (CCycleGAN) model to generate synthetic OD patterns from RD patterns and vice versa. For the RD-to-OD direction, we train a new recognizer of OD using the synthetic patterns. For the OD-to-RD direction, we utilize a pre-trained HME recognizer of RD to recognize the synthetic RD patterns generated from the OD patterns. Our experiments show that the CCycleGAN model performs significantly better than the CycleGAN model in terms of recognition accuracy. The recognition results demonstrate the effectiveness of our method in recognizing OD patterns using labeled RD patterns and unlabeled OD patterns for training. In addition, our CCycleGAN can well perform even when using only 100 training patterns in OD.
• 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 16, 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