DeepDocGen
🌟 Welcome to DeepDoc Generator World 🌟
This repository hosts all the code used to build our cutting-edge synthetic document pipeline. Dive into the world of synthetic document generation and object detection powered by deep learning!
📝 Paper Title
Generation and Detection of Objects in Documents by Deep Learning Neural Network Models (DeepDocGen)
📢 Our paper is coming soon! Stay tuned for the official release.
📖 About the Project
DeepDocGen is a pipeline designed to overcome the challenges of limited annotated datasets in document analysis. By generating synthetic documents, DeepDocGen empowers researchers and developers to train and fine-tune object detection models with high accuracy and robustness.
🛠️ Technologies Used
DeepDocGen leverages the powerful Mesh-Candidate Bestfit, an automated pipeline designed to synthesize large-scale, high-quality, and visually coherent document layout detection datasets.
In addition, DeepDocGen integrates the Content Generator, which intelligently fills bounding boxes with diverse types of content such as titles, text, tables, lists and more others.
This content placement ensures precise annotation, providing accurate knowledge of the exact position and type of each element within the generated layouts.
Depending on the label associated with each bounding box, the Content Generator applies specific formatting rules:
- Text boxes: the text can be aligned justified, right, left, or center. Rules such as starting sentences with a capital letter, maintaining consistent line spacing, and ending sentences with periods are respected.
- Title boxes: titles can start with a numbering (e.g., “1.”) followed by the title text, which can be either fully capitalized or start with a capital letter.
- List boxes: lists can begin with bullets like “*”, “-“, “•”, “[1]”, or “{1}”.
- Table boxes: tables are drawn and filled with appropriate content in each cell or column.
- Image boxes: a random image (JPG or PNG) is selected from the resource block and resized to fit the bounding box while maintaining visual consistency.
These strategies ensure that the synthetic documents generated by DeepDocGen maintain a realistic and structured layout, enhancing the quality of datasets for document layout detection tasks.
🎨 Visual Overview
🛠️ DeepDocGen Workflow
📊🎥 Some Document Sample Created
💡 How to Get Started
- Clone this repository:
git clone https://github.com/Geo99pro/DeepDocGen.git