HTR model for (Japanese) Kuzushiji

Training data came from the Kuronet Dataset of Japanese Classical Texts in Cursive Script (Owned by the National Institute of Japanese Literature and others/Processed by CODH) doi:10.20676/00000340 available at http://codh.rois.ac.jp/char-shape/ at the ROIS-DS Center for Open Data in the Humanities and created by Asanobu Kitamoto et alii. It was finetuned on top of the Chat model trained by Colin Brisson (https://zenodo.org/records/8383732) and has now 16906 codepoints, many of which are simply survivals from the Chat training and were not part of the Kuzushiji dataset. We modified the character bounding-box based annotation schema of Kuronet to linebased alto-XML necessary for kraken. The line polygons are convex-hulls of the character bounding-boxes according to basic heuristics. There is currently no segmentation inside kraken that would produce a similar segmentation. Funded by the European Union (ERC, MiDRASH, Project No. 101071829). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. This work has received funding through the Investissements d'Avenir program of the Agence Nationale de la Recherche with reference ANR-21-ESRE-0005 (Biblissima+).