File size: 2,083 Bytes
e98fa1d
 
 
 
 
 
 
 
 
 
 
 
 
 
462a6b5
5fdfb46
e98fa1d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-nc-4.0
task_categories:
- image-classification
language: [en, fr, es, tr, az, uz, ru, kk, ky, ar, zh]
size_categories:
- 1M<n<10M
tags:
- document-analysis
- font-recognition
- script-identification
- typography
models:
- issai/FontID
datasets:
- issai/kazparc
---

# DataFontID

A stratified synthetic corpus of **1,199,562** zero-margin text strips for multi-attribute typographic recognition. Every image carries four labels: font family, language, text color, and typographic style.

| | |
|---|---|
| Images | 1,199,562 (1,079,605 train / 59,978 val / 59,979 test) |
| Font families | 75 |
| Languages | 11 across Latin, Cyrillic, Arabic, and Han scripts |
| Colors | 64 (EGA palette) |
| Styles | Regular, Bold, Italic, Bold-Italic |
| Backgrounds | Texture, Noise, Complex, Document, Solid |
| Geometry | Zero-margin strips matching OCR extraction, variable aspect ratio |

Existing font corpora render text onto fixed square canvases. OCR engines emit tightly cropped strips with variable aspect ratios, so models trained on square canvases face a domain shift at inference. DataFontID renders into strips bounded by precise font metrics, matching the geometry produced at extraction time.

## Splits

The split is defined on the source pools before rendering, so no sentence and no background image seen during training appears in either evaluation split.

## Related resources

- Model: [issai/FontID](https://huggingface.co/issai/FontID)
- Real-world benchmark: [issai/Wild1024](https://huggingface.co/datasets/issai/Wild1024)

## Licensing

The DataFontID annotations, rendering pipeline, and synthesized text are released
under CC BY-NC 4.0. Background imagery for the Texture, Noise, and Complex
categories derives from the Describable Textures Dataset, which its authors make
available to the computer vision community for research purposes; users should
observe those terms for the background content. Text is drawn from KazParC and
CulturaX. The corpus contains rendered raster images only and no font binaries.

## Citation

Paper is coming soon.