Synthetic Text Image Generator for OCR Model | Paper | Documentation | Datasets
The documentation is available at https://clovaai.github.io/synthtiger/.
You can check API reference in this documentation.
SynthTIGER requires python>=3.6 and libraqm.
To install SynthTIGER from PyPI:
$ pip install synthtigerIf you see a dependency error when you install or run SynthTIGER, install dependencies.
# Set environment variable (for macOS)
$ export OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YESusage: synthtiger [-h] [-o DIR] [-c NUM] [-w NUM] [-s NUM] [-v] SCRIPT NAME [CONFIG]
positional arguments:
SCRIPT Script file path.
NAME Template class name.
CONFIG Config file path.
optional arguments:
-h, --help show this help message and exit
-o DIR, --output DIR Directory path to save data.
-c NUM, --count NUM Number of output data. [default: 100]
-w NUM, --worker NUM Number of workers. If 0, It generates data in the main process. [default: 0]
-s NUM, --seed NUM Random seed. [default: None]
-v, --verbose Print error messages while generating data.
# horizontal
synthtiger -o results -w 4 -v examples/synthtiger/template.py SynthTiger examples/synthtiger/config_horizontal.yaml
# vertical
synthtiger -o results -w 4 -v examples/synthtiger/template.py SynthTiger examples/synthtiger/config_vertical.yaml
images: a directory containing images.gt.txt: a file containing text labels.coords.txt: a file containing bounding boxes of characters with text effect.glyph_coords.txt: a file containing bounding boxes of characters without text effect.masks: a directory containing mask images with text effect.glyph_masks: a directory containing mask images without text effect.
synthtiger -o results -w 4 -v examples/multiline/template.py Multiline examples/multiline/config.yaml
images: a directory containing images.gt.txt: a file containing text labels.
-
Prepare corpus
txtformat, line by line (example). -
Prepare fonts
See font customization for more details.
-
Edit corpus path and font path in config file (example)
-
Run synthtiger
-
Prepare fonts
ttf/otfformat (example). -
Extract renderable charsets
python tools/extract_font_charset.py -w 4 fonts/
This script extracts renderable charsets for all font files (example).
Text files are generated in the input path with the same names as the fonts.
-
Edit font path in config file (example)
-
Run synthtiger
-
Prepare images
jpg/jpeg/png/bmpformat. -
Create colormaps
python tools/create_colormap.py --max_k 3 -w 4 images/ colormap.txt
This script creates colormaps for all image files (example).
-
Edit colormap path in config file (example)
-
Run synthtiger
You can implement custom templates by inheriting the base template.
from synthtiger import templates
class MyTemplate(templates.Template):
def __init__(self, config=None):
# initialize template.
def generate(self):
# generate data.
def init_save(self, root):
# initialize something before save.
def save(self, root, data, idx):
# save data to specific path.
def end_save(self, root):
# finalize something after save.SynthTIGER is available for download at google drive.
Dataset was split into several smaller files. Please download all files and run following command.
# for Linux, macOS
cat synthtiger_v1.0.zip.* > synthtiger_v1.0.zip
# for Windows
copy /b synthtiger_v1.0.zip.* synthtiger_v1.0.zipsynthtiger_v1.0.zip (36G) (md5: 5b5365f4fe15de24e403a9256079be70)
- Original paper version.
- Used MJ and ST label.
synthtiger_v1.1.zip (38G) (md5: b2757a7e2b5040b14ed64c473533b592)
- Used MJ and ST lexicon instead of MJ and ST label.
- Fixed a bug that applies transformation twice on curved text.
- Fixed a bug that incorrectly converts grayscale to RGB.
| Version | IIIT5k | SVT | IC03 | IC13 | IC15 | SVTP | CUTE80 | Total |
|---|---|---|---|---|---|---|---|---|
| 1.0 | 93.2 | 87.3 | 90.5 | 92.9 | 72.1 | 77.7 | 80.6 | 85.9 |
| 1.1 | 93.4 | 87.6 | 91.4 | 93.2 | 73.9 | 77.8 | 80.6 | 86.6 |
The structure of the dataset is as follows. The dataset contains 10M images.
gt.txt
images/
0/
0.jpg
1.jpg
...
9998.jpg
9999.jpg
1/
...
998/
999/
The format of gt.txt is as follows. Image path and label are separated by tab. (<image_path>\t<label>)
images/0/0.jpg 10
images/0/1.jpg date:
...
images/999/9999998.jpg STUFFIER
images/999/9999999.jpg Re:
@inproceedings{yim2021synthtiger,
title={SynthTIGER: Synthetic Text Image GEneratoR Towards Better Text Recognition Models},
author={Yim, Moonbin and Kim, Yoonsik and Cho, Han-Cheol and Park, Sungrae},
booktitle={International Conference on Document Analysis and Recognition},
pages={109--124},
year={2021},
organization={Springer}
}SynthTIGER
Copyright (c) 2021-present NAVER Corp.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
The following directories and their subdirectories are licensed the same as their origins. Please refer to NOTICE
docs/
resources/font/
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent
