File size: 1,664 Bytes
71966ed
2b89df8
71966ed
2b89df8
 
 
 
d76be98
 
 
 
2b89df8
 
 
 
 
 
 
46969a6
71966ed
2b89df8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f2b942
2b89df8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---
library_name: PyLaia
license: mit
tags:
- PyLaia
- PyTorch
- Handwritten text recognition
- Automatic text recognition
- HTR
- OCR
- ATR
metrics:
- CER
- WER
language:
- 'fr'
datasets:
- Teklia/Belfort
pipeline_tag: image-to-text
---

# Belfort handwritten text recognition

This model performs Handwritten Text Recognition in French on historical documents.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](https://zenodo.org/records/8041668).

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

| split | N 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 (%) | N lines   |
|:------|:---------------|:----------:|:-------:|----------:|
| test  | no             | 10.54      |   28.12 |     3,819 |
| test  | yes            |  9.52      |   23.73 |     3,819 |

## How to use

Please refer to the [documentation](https://atr.pages.teklia.com/pylaia/).

## Cite us

```bibtex
@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"
}
```