DehazeGAN: Sand-Dust Image Restoration using Semi-supervised Adversarial Learning

About The Project

Removing color cast for image restoration in sand and dust conditions is essential for improving the accuracy and performance of outdoor computer vision systems. Enhancements in this area are crucial for environments heavily impacted by sand and dust, where visibility and image quality are significantly degraded. Despite so, only few deep learning models are published to denoise and remove color cast from sandstorm images.

Therefore, we propose a novel deep-learning model extended from TOENet (Gao et al., 2023), DehazeGAN to remove color cast from sandstorm images through unsupervised fine-tuning (UFT) using adversarial learning and supervised fine-tuning (SFT) via paired augmentation.

Getting Started

Installation Prerequisites

We developed and tested our application using the following configurations.

• OS: Linux Ubuntu 22.04
• Python Version: 3.10
• Nvidia Driver Version: 545.29.06
• Nvidia CUDA Tool-kit Version: 12.0.1

Currently, our project only supports device with GPU. If your device does not have GPU, please consider using cloud services or remote desktop with GPU.

Installation Guide

In the terminal, clone a repository for our project from GitHub using git command

git clone https://github.com/superleesa/final_year_project.git

2. Go to the repository directory and install Python libraries required for the app, using the command

pip install -r requirements.txt.

Training Guidelines

For training of the dataset, we would need to download the SIE Dataset. The paper on the SIE Dataset can be found here: LINK. The SIE Dataset should be downloaded to the Data folder, in the format of the Data DIrectory Structure below.

Data Directory Structure

Data/
└── paired/
|   ├── ground_truth/
|   │   ├── ...
|   │   ├── ...
|   │   └── ...
|   └── noisy/
|       ├── ...
|       ├── ...
|       └── ...
└── unpaired/
    ├── clear/
    │   ├── ...
    │   ├── ...
    │   └── ...
    └── noisy/
        ├── ...
        ├── ...
        └── ...

Training of Model

We follow the training pipeline as illustrated below:

• Unsupervised Finetuning Training (UFT): Refer to README.md for UFT Training.
• Supervised Finetuning Training (SFT): Refer to README.md for SFT Training.

Evaluation Guidelines

See README.md for Evaluation Guidelines.

Results

Metrics Comparison with Other Models

Image Color Histogram Comparison with Other Models

Paired Synthetic Image Comparison with Other Models

Unpaired Real-world Image Comparison with Other Models

Acknowledgements

The base model of our project is based on the [TOENet] (https://github.com/YuanGao-YG/TOENet).