Applying Handwritten Text Recognition to the Distinctiones Collections
Résumé
Due to the development of neural network-based techniques, Handwritten Text Recognition is undergoing an unprecedented rise. A few years ago, the dominant approach was to create a unique model for each new project. Today, the more popular strategy is transfer learning followed by fine-tuning, made possible by initiatives such as HTR-united, CremmaLab, and others.
The DISTINGUO project aims at the text extraction of the collections of distinctiones from scanned manuscripts and incunables and their publication in the form of a searchable database. Given the massive variety of hands and fonts in the corpus and limited human resources, we decided to use eScriptorium to create one general model, which is fine-tuned for each particular manuscript or book. However, this approach raises the question of how much data is needed so that the model could serve as a pre-trained base to raise a new specialized model.
This presentation aims at suggesting the answer to this question and focusing on calculating the learning break point of the fine-tuned models, as well as on the post-processing techniques that can improve the quality of the transcribed text.
Domaines
Histoire
Loading...