metadata

library_name: PyLaia
license: mit
tags:
  - PyLaia
  - PyTorch
  - atr
  - htr
  - ocr
  - historical
  - handwritten
metrics:
  - CER
  - WER
language:
  - fr
datasets:
  - Teklia/POPP
pipeline_tag: image-to-text

PyLaia - POPP

This model performs Handwritten Text Recognition on French census documents.

Model description

The model was trained using the PyLaia library on the POPP generic.

For training, text-lines were resized with a fixed height of 128 pixels, keeping the original aspect ratio.

split	N lines
train	3,835
val	480
test	479

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

Evaluation results

The model achieves the following results:

set	Language model	CER (%)	WER (%)	N lines
test	no	16.49	36.26	479
test	yes	16.09	34.52	479

How to use

Please refer to the documentation.

Cite us

@inproceedings{pylaia-lib,
    author = "Tarride, Solène and Schneider, Yoann and Generali, Marie and Boillet, Melodie and Abadie, Bastien and Kermorvant, Christopher",
    title = "Improving Automatic Text Recognition with Language Models in the PyLaia Open-Source Library",
    booktitle = "Submitted at ICDAR2024",
    year = "2024"
}