TrOCR Kurrent-Model 16th to 18th century

Handwritten Text Recognition model for 16th to 18th century German.

Part of the developments at the Digital Humanities@University of Bern. Developed by Jonas Widmer, Anna Funk, Christa Schneider, and Tobias Hodel.

Base model: dh-unibe/trocr-kurrent

Epochs: 19.85 / 20
Eval CER: 0.05673
Test CER: 0.05416

This model is based on an extensive training set (of roughly 1579200 words) and evaluated against the same hands in an evaluation and test set (automatic split). Consisting of German Kurrent scripts written in the 16th-18th century.

The ground truth stems from different projects and partners and is biased toward Swiss documents. It is based on documents from a variety of archives and projects. Among others, the State Archives of Zürich (Stillstandsprotokolle, Ratsmanuale, Findmittel), and the scholarly edition project Königsfelden (Universitäten Zürich und Bern: www.koenigsfelden.uzh.ch). As well as transcriptions from Einsiedeln. Further contributions by the university archives of Greifswald: https://rechtsprechung-im-ostseeraum.archiv.uni-greifswald.de/.

The public Transkribus model (based on PyLaia) can be found here: https://readcoop.eu/model/german-kurrent-16th-18th/

Extensive testing of the model has still to be carried out. This is only a first attempt but might help for fine-tuning tasks.

Input: one line, not a page

This is a VisionEncoderDecoderModel for line images: it reads one cropped text line per call, which is also what the CER above was measured on. A whole page is not an input it has — pass one and it returns a fragment of it.

Segment the page first (kraken, eScriptorium, Transkribus, or any line segmenter) and pass the crops one at a time. In the DH Bern serving stack the model is registered level: line and the gateway segments with kraken before calling it; the reasoning and the measurements behind that rule are in thodel/serving-atr-inference#165.