TariMa models

Open version of OCR/HTR models trained with the Tarima dataset, gathered within the frame of the Tarima project (Tarih al-Maghrib. Writing History in the Maghreb in the modern and contemporary era), sponsored by the French agency Collex-Persee and supervised by Antoine Perrier (CNRS).

The project combines the development of character recognition models for manuscripts (HTR) and prints (OCR) of Maghrebi Arabic texts and the scientific study of collections of Maghrebi works held in French libraries (BULAC, project manager, College de France and Bibliothèque nationale universitaire de Strasbourg).

It aims to make available the digitization in image and text format of a set of primary sources, that have been little used until now. Making them available online will promote the use of Arabic sources, which are at the heart of recent shifts in the historiography of the Maghreb.

The BULAC library releases three open models trained with kraken architecture (default settings), one dedicated to lithographies, one to manuscripts and one to layout analysis. These models have not been used during the project, but achieved a mean accuracy of 85% and can be used as a basis to fine-tune new models for Maghrebi Arabic scripts.

Settings, accuracy, classes and data split can be found in the .mlmodel files. See the CoreMLTools library for more details : https://coremltools.readme.io/docs

To know more about the dataset : https://github.com/calfa-co/tarima

To know more about the annotation tool : https://vision.calfa.fr

To know more about how to use models : http://kraken.re/

Related publications to this project :

• Vidal-Gorène, C., Lucas, N., Salah, C., Decours-Perez, A., & Dupin, B. (2021, September). RASAM–A Dataset for the Recognition and Analysis of Scripts in Arabic Maghrebi. In Document Analysis and Recognition–ICDAR 2021 Workshops: Lausanne, Switzerland, September 5–10, 2021, Proceedings, Part I (pp. 265-281). Cham: Springer International Publishing. see on HAL

• Lucas, N., Salah, C., & Vidal-Gorène, C. (2022). New Results for the Text Recognition of Arabic Maghribi Manuscripts - Managing an Under-resourced Script. arXiv preprint arXiv:2211.16147. see on HAL

This work was carried out with the financial support of the French agency Collex-Persée, and gathered four institutions: the BULAC library, the Maison Méditerranéenne des Sciences de l'Homme, the IREMAM institute and Calfa.