Text Renderer

Generate text line images for training deep learning OCR model (e.g. CRNN).

• Modular design. You can easily add different components: Corpus, Effect, Layout.
• Integrate with imgaug, see imgaug_example for usage.
• Support render multi corpus on image with different effects. Layout is responsible for the layout between multiple corpora
• Support apply effects on different stages of rendering process corpus_effects, layout_effects, render_effects.
• Generate vertical text.
• Support generate lmdb dataset which compatible with PaddleOCR, see Dataset
• A web font viewer.
• Corpus sampler: helpful to perform character balance

Documentation

Run Example

Run following command to generate images using example data:

git clone https://github.com/oh-my-ocr/text_renderer
cd text_renderer
python3 setup.py develop
pip3 install -r docker/requirements.txt
python3 main.py \
    --config example_data/example.py \
    --dataset img \
    --num_processes 2 \
    --log_period 10

The data is generated in the example_data/output directory. A labels.json file contains all annotations in follow format:

{
  "labels": {
    "000000000": "test",
    "000000001": "text2"
  },
  "sizes": {
    "000000000": [
      120,
      32 
    ],
    "000000001": [
      128,
      32 
    ]
  },
  "num-samples": 2
}

You can also use --dataset lmdb to store image in lmdb file, lmdb file contains follow keys:

• num-samples
• image-000000000
• label-000000000
• size-000000000

You can check config file example_data/example.py to learn how to use text_renderer, or follow the Quick Start to learn how to setup configuration

Quick Start

Prepare file resources

• Font files: .ttf、.otf、.ttc
• Background images of any size, either from your business scenario or from publicly available datasets (COCO, VOC)
• Corpus: text_renderer offers a wide variety of text sampling methods, to use these methods, you need to consider the preparation of the corpus from two perspectives：

1. The corpus must be in the target language for which you want to perform OCR recognition
2. The corpus should meets your actual business needs, such as education field, medical field, etc.

• Charset file [Optional but recommend]: OCR models in real-world scenarios (e.g. CRNN) usually support only a limited character set, so it's better to filter out characters outside the character set during data generation. You can do this by setting the chars_file parameter

You can download pre-prepared file resources for this Quick Start from here:

• simsun.ttf
• background.png
• eng_text.txt

Save these resource files in the same directory:

workspace
├── bg
│ └── background.png
├── corpus
│ └── eng_text.txt
└── font
    └── simsun.ttf

Create config file

Create a config.py file in workspace directory. One configuration file must have a configs variable, it's a list of GeneratorCfg.

The complete configuration file is as follows:

import os
from pathlib import Path

from text_renderer.effect import *
from text_renderer.corpus import *
from text_renderer.config import (
    RenderCfg,
    NormPerspectiveTransformCfg,
    GeneratorCfg,
    SimpleTextColorCfg,
)

CURRENT_DIR = Path(os.path.abspath(os.path.dirname(__file__)))


def story_data():
    return GeneratorCfg(
        num_image=10,
        save_dir=CURRENT_DIR / "output",
        render_cfg=RenderCfg(
            bg_dir=CURRENT_DIR / "bg",
            height=32,
            perspective_transform=NormPerspectiveTransformCfg(20, 20, 1.5),
            corpus=WordCorpus(
                WordCorpusCfg(
                    text_paths=[CURRENT_DIR / "corpus" / "eng_text.txt"],
                    font_dir=CURRENT_DIR / "font",
                    font_size=(20, 30),
                    num_word=(2, 3),
                ),
            ),
            corpus_effects=Effects(Line(0.9, thickness=(2, 5))),
            gray=False,
            text_color_cfg=SimpleTextColorCfg(),
        ),
    )


configs = [story_data()]

In the above configuration we have done the following things:

1. Specify the location of the resource file
2. Specified text sampling method: 2 or 3 words are randomly selected from the corpus
3. Configured some effects for generation
   • Perspective transformation NormPerspectiveTransformCfg
   • Random Line Effect
   • Fix output image height to 32
   • Generate color image. gray=False, SimpleTextColorCfg()
4. Specifies font-related parameters: font_size, font_dir

Run

Run main.py, it only has 4 arguments:

• config：Python config file path
• dataset: Dataset format img or lmdb
• num_processes: Number of processes used
• log_period: Period of log printing. (0, 100)

All Effect/Layout Examples

Find all effect/layout config example at link

• bg_and_text_mask: Three images of the same width are merged together horizontally, it can be used to train GAN model like EraseNet

	Name	Example
0	bg_and_text_mask
1	char_spacing_compact
2	char_spacing_large
3	color_image
4	curve
5	dropout_horizontal
6	dropout_rand
7	dropout_vertical
8	emboss
9	extra_text_line_layout
10	line_bottom
11	line_bottom_left
12	line_bottom_right
13	line_horizontal_middle
14	line_left
15	line_right
16	line_top
17	line_top_left
18	line_top_right
19	line_vertical_middle
20	padding
21	perspective_transform
22	same_line_layout_different_font_size
23	vertical_text

Contribution

Setup Commitizen for commit message

• Corpus: Feel free to contribute more corpus generators to the project, It does not necessarily need to be a generic corpus generator, but can also be a business-specific generator, such as generating ID numbers

Run in Docker

Build image

docker build -f docker/Dockerfile -t text_renderer .

Config file is provided by CONFIG environment. In example.py file, data is generated in example_data/output directory, so we map this directory to the host.

docker run --rm \
-v `pwd`/example_data/docker_output/:/app/example_data/output \
--env CONFIG=/app/example_data/example.py \
--env DATASET=img \
--env NUM_PROCESSES=2 \
--env LOG_PERIOD=10 \
text_renderer

Font Viewer

Start font viewer

streamlit run tools/font_viewer.py -- web /path/to/fonts_dir

Build docs

cd docs
make html
open _build/html/index.html

Citing text_renderer

If you use text_renderer in your research, please consider use the following BibTeX entry.

@misc{text_renderer,
  author =       {oh-my-ocr},
  title =        {text_renderer},
  howpublished = {\url{https://github.com/oh-my-ocr/text_renderer}},
  year =         {2021}
}