The Reproduction and Quality Evaluation of Image Dehazing Algorithms

Introduction

This repository contains the Python code of reproducton of some image dehazing algorithms (DHA), including:

• DCP (Dark Channel Prior Dehazing Method)
• FVR (Fast Visibility Restoration Method)
• AOD-Net (All-in-One Dehazing Network Method)

For each dehazing method there is a Python file to implement the algorithm.

Usage

Dataset Directory

To run this program correctly, you need to create several folders to store the datasets. In the root of the project directory, create the following directories:

.
├── AOD-net-snapshots
├── samples
├── test-data-aod
│   ├── dehazed
│   ├── evaluate
│   ├── GT
│   └── hazy
├── test-data-dcp
│   ├── dark-channel-prior
│   ├── dehazed
│   ├── evaluate
│   ├── GT
│   └── hazy
├── test-data-fvr
│   ├── dehazed
│   ├── evaluate
│   ├── GT
│   ├── hazy
│   ├── veil
│   └── white-balanced
├── test-data-hist
│   ├── dehazed
│   ├── evaluate
│   ├── GT
│   └── hazy
└── training-image-AOD-net
    ├── data
    └── images

Put the hazy or not dehazed image in the hazy/ directory, while put the clear image (also known as "GT") in GT/ directory if you have prepared the images for evaluation. Specifically for the AOD-Net training, put the original clear image in training-image-AOD-net/images/ directory, and put the hazy image in training-image-AOD-net/data/ directory. After the training of AOD-Net, the model will be stored in AOD-net-snapshots directory.

Environment Configuration

We use uv to manage this project, so we recommend you to install a copy of uv in your system to handle the Python projects. Here are the steps.

Install the proper version of Python

uv python install 3.13

Create a virtual environment

If you use PyCharm, you can simply create a virtual environment for this project in GUI through uv; if you use other IDEs, you can also use uv command line to create a virtual environment.

This project is based on Python 3.13, so just install the specific verison if you don't have Python 3.13 installed in your computer.

uv venv --python 3.13

Install the requirements

Auto install the requirements in your IDE, or install manually:

uv pip install [package_name]

Start training AOD-Net

Run the AOD_train.py to train the AOD-Net model for image dehazing. Click the "run single file" button or:

uv run AOD_train.py

This process may take a little long time. The program will exit automatically when the training task is finished.

Test dehazing methods

Run single file or run all methods at a time, the evaluation result will be stored in evaluate/ directory as csv files.