FairScan Segmentation Model

This repository contains a lightweight, custom-trained model for document segmentation. Its purpose is to provide FairScan with automatic document detection.

Overview

• Model architecture: DeepLabV3Plus with MobileNet v2 encoder
• Format: PyTorch (.pt) and quantized TFLite (.tflite)
• Use case: Detecting the area of a document in a photo, as a first step in document scanning workflows
• Target platform: Android (LiteRT / TensorFlow Lite)

Features

• Small model size for fast inference on mobile
• High segmentation accuracy (Dice score > 0.94 on validation set)
• Supports quantization for optimized deployment
• Compatible with LiteRT
• Easily reproducible training: run training with a single command

Dataset

The dataset can be found in a separate repository: fairscan-segmentation-dataset. It's automatically downloaded in the training script.

Training

# 1. Clone the repository
git clone https://github.com/pynicolas/fairscan-segmentation-model
cd fairscan-segmentation-model

# 2. Create a venv
python3 -m venv venv
source venv/bin/activate  # or venv\Scripts\activate.bat on Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. Run the training script
python train.py

Requirements

See requirements.txt for details.

License

This repository is released under the GNU GPLv3 license. See LICENSE for details.

Acknowledgements

This project builds on the work of several excellent open-source projects:

• PyTorch — Deep learning framework (BSD-3-Clause)
• Albumentations — Image augmentation library (MIT)
• segmentation-models-pytorch — High-level segmentation models including U-Net (MIT)
• OpenCV — Image processing library (Apache-2.0)
• scikit-learn — Machine learning utilities (BSD-3-Clause)
• ai-edge-torch — Model quantization and export to TFLite (Apache-2.0)

We gratefully acknowledge the work of the authors and communities behind these libraries.