DoGe — Synthetic DOcument GEnerator for Document AI

DoGe is designed to synthesize a dataset of realistic document scans. Each document contains meaningful text, headings, tables, paragraphs with different formatting and fonts which are parsed from Wikipedia. The coordinates of the words are extracted using the No-OCR method we invented for faster generation on CPU.

Document examples

Check the full size (1024x1024) in resources folder.

Usage

Docker

You can use Docker image with predefined environment to run DoGe:

git clone https://github.com/Travvy88/DocumentGenerator_DoGe
cd DocumentGenerator_DoGe
docker build -t doge .
docker run -it -v /path/to/output/folder/on/host:/app/data doge

Replace /path/to/output/folder/on/host and run commands. Inside the docker container you can start document generation.

Ubuntu

For faster generation, it is recommended to install all dependencies without Docker. Doge is tested on Ubuntu 22.04.

sudo apt-get update && apt-get install libreoffice libjpeg-dev zlib1g-dev poppler-utils
/usr/bin/python3 -m pip install --user unoserver  # install unoserver to system python

git clone https://github.com/Travvy88/DocumentGenerator_DoGe
cd DocumentGenerator_DoGe
pip3 install -r requirements.txt  # there you can make venv if needed!

Start Data Generation

python3 main.py --out_dir data --image_size 244 --max_urls 4 --num_processes 2 --ports 4000 4001 4002 4003

Main.py

The following arguments can be passed to the script:

• --out_dir: The output directory for saving results. This argument is required.
• --remove_existing_dir: If set to True, the output directory will be deleted before creating a new one. Default is False.
• --image_size: The size of the final images. Default is 244.
• --start_page: The starting page URL. Default is the Wikipedia main English page. You can use another language Wiki man page URL.
• --languages: Permitted languages. Pages with other localizations will be ignored. Default is ['en'].
• --max_urls: The maximum number of URLs to process. Default is 100.
• --num_processes: The number of processes to use. Default is 1. Each process will start DocumentGenerator inside and start separate Unoserver.
• --max_threads: The maximum threads inside a process. Default is 3.
• --ports: The list of ports to use. Default is [8145, 8146]. The number of ports should be 2 times larger than num_processes.

Docx_config.json

Parameter	Description
max_words	The maximum number of words allowed in the generated documents.
p_2columns	The probability that the document will be formatted into two columns.
font_size_interval	The font size range from which the size is randomly selected for each document.
p_line_spacing	A list of probabilities controlling the line spacing of the document (1.5 or double).
p_text_alignment	A list of probabilities controlling the text alignment of the document (center, left, right, justify).
p_heading_bold	The probability that headings will be displayed in bold font.
heading_relative_size_interval	The range of relative font sizes for headings. The relative font size is chosen randomly.
p_heading_alignment	A list of probabilities controlling the alignment of headings (center, left, right, justify).
table_max_rows	The maximum number of rows allowed in a table. Tables with more than the specified number of rows are dropped.
table_max_cols	The maximum number of columns allowed in a table. Tables with more than the specified number of columns are dropped.

Parameters with probabilities and intervals calculate its values for each document randomly.

According to my experience, generator produces an average about 14 images for each url with the above Docx settings.

Augmentations

Augmentation pipeline applies on a final stage. You can manage different augmentations in src/augmentations.py file. Read the Augraphy Docs for detailed explanation.

How it works

Firstly, the Manager class creates the DocumentGenerator instances in separate processes. For each DocumentGenerator, a Unoserver instance is started.

Then, the UrlParser generates a list of URLs by crawling the web, starting from a given start page and following links on each page. It uses BeautifulSoup to parse HTML content and extract links, then checks each link's validity and language, adding it to the list if it meets certain conditions. The process continues until a maximum number of URLs is reached, and the method returns the list of generated URLs, excluding the starting URL.

When data generation begins, the list of URLs is divided into several chunks for each DocumentGenerator. Each DocumentGenerator instance retrieves a Wikipedia HTML page by URL from its chunk. Headers, paragraphs formatting, and tables are extracted and placed into a Docx document via the DocxDocument class. At this stage, some random parametrization is applied according to docx_config.json. For example, font size, text alignment, one or two columns, and other parameters are chosen for each document randomly.

After that, each word in the Docx is filled with a unique color. As a result, a colored rectangle appears in place of each word. The image will be encoded with 24-bit color depth, so the maximum number of words per document is 16,777,216. The text of each word is saved to a hashmap of type color_code -> word.

The next step is Docx to image conversion. DoGe uses Unoserver to convert Docx to Pdf and pdf2image for image rendering.

Then, all rectangle coordinates are detected via OpenCV on converted images. The word for each bounding box is retrieved from the hashmap. DoGe saves annotations to JSON files in the following format:

{
  "words": [
    "Hello", 
    "World"
  ],
  "bboxes": [
    [0.1, 0.1, 0.03, 0.02],
    [0.4, 0.3, 0.11, 0.02]
  ]
}

The bboxes are normalized and saved in XYWH format.

The final step is deleting all color fills from words in the Docx document, rendering images, applying Augraphy augmentations, and saving the augmented images to disk. That's it!

Future Plan

• Improve pipeline to download and place images into documents
• Add annotations of headers, tables, paragraphs
• Add different output formats (Parquet for example)

Acknowledgments

Here are some great open-source projects I benefit from:

• ISP RAS Dedoc Team for support and assistance.
• Augraphy for augmentation code of final images.
• Unoserver for Docx to Pdf converter.
• Pdf2image for image from Pdf rendering module.
• Pillow-SIMD for faster image processing.