Teklia/pylaia-belfort

PyLaia - Belfort

This model performs Handwritten Text Recognition in French on historical documents.

Model description

The model was trained using the PyLaia library on the Belfort dataset.

Training images were resized with a fixed height of {dimension} pixels, keeping the original aspect ratio. Vertical lines are discarded.

set	lines
train	25,800
val	3,102
test	3,819

An external 6-gram character language model can be used to improve recognition. The language model is trained on the text from the Belfort training set.

Evaluation results

The model achieves the following results:

set	Language model	CER (%)	WER (%)	lines
test	no	10.54	28.12	3,819
test	yes	9.52	23.73	3,819

How to use?

Please refer to the PyLaia documentation to use this model.

Cite us!

@inproceedings{pylaia2024,
    author = {Tarride, Solène and Schneider, Yoann and Generali-Lince, Marie and Boillet, Mélodie and Abadie, Bastien and Kermorvant, Christopher},
    title = {{Improving Automatic Text Recognition with Language Models in the PyLaia Open-Source Library}},
    booktitle = {Document Analysis and Recognition - ICDAR 2024},
    year = {2024},
    publisher = {Springer Nature Switzerland},
    address = {Cham},
    pages = {387--404},
    isbn = {978-3-031-70549-6}
}

@inproceedings{belfort-2023,
    author = {Tarride, Solène and Faine, Tristan and Boillet, Mélodie and Mouchère, Harold and Kermorvant, Christopher},
    title = {Handwritten Text Recognition from Crowdsourced Annotations},
    year = {2023},
    isbn = {9798400708411},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3604951.3605517},
    doi = {10.1145/3604951.3605517},
    booktitle = {Proceedings of the 7th International Workshop on Historical Document Imaging and Processing},
    pages = {1–6},
    numpages = {6},
    keywords = {Crowdsourcing, Handwritten Text Recognition, Historical Documents, Neural Networks, Text Aggregation},
    series = {HIP '23}
}