{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/boosting-handwriting-text-recognition-in","title":"Boosting Handwriting Text Recognition in Small Databases with Transfer Learning","arxiv_id":"1804.01527","date":"2018-04-04","proceeding":null,"authors":["José Carlos Aradillas","Juan José Murillo-Fuentes","Pablo M. Olmos"],"abstract":"In this paper we deal with the offline handwriting text recognition (HTR)\nproblem with reduced training datasets. Recent HTR solutions based on\nartificial neural networks exhibit remarkable solutions in referenced\ndatabases. These deep learning neural networks are composed of both\nconvolutional (CNN) and long short-term memory recurrent units (LSTM). In\naddition, connectionist temporal classification (CTC) is the key to avoid\nsegmentation at character level, greatly facilitating the labeling task. One of\nthe main drawbacks of the CNNLSTM-CTC (CLC) solutions is that they need a\nconsiderable part of the text to be transcribed for every type of calligraphy,\ntypically in the order of a few thousands of lines. Furthermore, in some\nscenarios the text to transcribe is not that long, e.g. in the Washington\ndatabase. The CLC typically overfits for this reduced number of training\nsamples. Our proposal is based on the transfer learning (TL) from the\nparameters learned with a bigger database. We first investigate, for a reduced\nand fixed number of training samples, 350 lines, how the learning from a large\ndatabase, the IAM, can be transferred to the learning of the CLC of a reduced\ndatabase, Washington. We focus on which layers of the network could be not\nre-trained. We conclude that the best solution is to re-train the whole CLC\nparameters initialized to the values obtained after the training of the CLC\nfrom the larger database. We also investigate results when the training size is\nfurther reduced. The differences in the CER are more remarkable when training\nwith just 350 lines, a CER of 3.3% is achieved with TL while we have a CER of\n18.2% when training from scratch. As a byproduct, the learning times are quite\nreduced. Similar good results are obtained from the Parzival database when\ntrained with this reduced number of lines and this new approach.","url_abs":"http://arxiv.org/abs/1804.01527v1","url_pdf":"http://arxiv.org/pdf/1804.01527v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"boosting-handwriting-text-recognition-in","repo_url":"https://github.com/josarajar/HTRTF","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":null},{"paper_slug":"boosting-handwriting-text-recognition-in","repo_url":"https://github.com/SamuelNguyen1998/Vietnamese_Handwriting_Recognition","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"MIT"}},{"paper_slug":"boosting-handwriting-text-recognition-in","repo_url":"https://github.com/huyhoang17/Vietnamese_Handwriting_Recognition","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"htr","task_name":"HTR"},{"task_slug":"transfer-learning","task_name":"Transfer Learning"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}