The previous project was done during the 2020 SURF program and focused on modifying and training Harald Scheidl’s HTR model [2] for line based Syriac texts. This project shifts its focus to Jonathan Chung and Thomas Delteil’s HTR model [1], aiming to achieve more accurate recognition. This project’s training was done in four steps: extract images and transcripts, load data with the necessary information, training, and evaluation. Methods like image augmentation are used to enlarge the dataset. The overall result is promising. However, there are issues with the proxies system, and improvements can be made in other parts of the framework as well. Thus, more research is needed to eliminate current problems and further improve the result. A poster derived from academic research with Nicholas Howe, Professor of Computer Science.
by Winnie Zong '23, Annie Portoghese '24 and Sophie Gutierrez '24