Computer Vision News - January 2023

21 Mohamed Ali Souibgui For text recognition in low- resource, the goal was to propose generic approaches to recognize different types of historical ciphered texts, as the ones shown in Fig. 2. First, A few-shot learning approach was proposed to recognize the historical ciphered manuscripts requiring only a few examples (usually 5) from any new script, even not seen during training. Second, a one-shot character generation method was used to generate handwritten text line images for training. The generation was requiring a single example of each character. After that, a self- supervised learning approach was also used to learn rich representations from unlabeled data. In the thesis, it was shown that these representations serve as a good start to train models, especially when the labeled training data is limited. Thus, it is a good solution for the low resource scenario when there is a big amount of data, but most of it is unlabeled. In the future, Mohamed plans to extend his work by exploring diffusion models for image enhancement and continual learning for HTR in low-resource. He will also focus on model robustness and explainability . Fig. 1: Qualitative results of enhancing the quality of the degraded images by our models. Fig. 2: Examples of handwritten ciphers dated from the 16th to the 18th century. | Top: Devil cipher Middle: Borg Cipher. Bottom: Copiale cipher.