FFA-Net: Feature Fusion Attention Network for Single Image Dehazing (AAAI 2020)

Official implementation.

---

by Xu Qin, Zhilin Wang et al. Peking University and Beijing University of Aeronautics & Astronautics.

Citation

@inproceedings{qin2020ffa,
title={FFA-Net: Feature fusion attention network for single image dehazing},
author={Qin, Xu and Wang, Zhilin and Bai, Yuanchao and Xie, Xiaodong and Jia, Huizhu},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={34},
number={07},
pages={11908--11915},
year={2020}
}

Dependencies and Installation

• python3
• PyTorch>=1.0
• NVIDIA GPU+CUDA
• numpy
• matplotlib
• tensorboardX(optional)

Datasets Preparation

Dataset website:RESIDE ; Paper arXiv version:[RESIDE: A Benchmark for Single Image Dehazing]

FILE STRUCTURE

    FFA-Net
    |-- README.md
    |-- net
    |-- data
        |-- RESIDE
            |-- ITS
                |-- hazy
                    |-- *.png
                |-- clear
                    |-- *.png
            |-- OTS 
                |-- hazy
                    |-- *.jpg
                |-- clear
                    |-- *.jpg
            |-- SOTS
                |-- indoor
                    |-- hazy
                        |-- *.png
                    |-- clear
                        |-- *.png
                |-- outdoor
                    |-- hazy
                        |-- *.jpg
                    |-- clear
                        |-- *.png

Metrics update

Methods	Indoor(PSNR/SSIM)	Outdoor(PSNR/SSIM)
DCP	16.62/0.8179	19.13/0.8148
AOD-Net	19.06/0.8504	20.29/0.8765
DehazeNet	21.14/0.8472	22.46/0.8514
GFN	22.30/0.8800	21.55/0.8444
GCANet	30.23/0.9800	-/-
Ours	36.39/0.9886	33.57/0.9840

Usage

Train

Remove annotation from main.py if you want to use tensorboard or view intermediate predictions

If you have more computing resources, expanding bs, crop_size, gps, blocks will lead to better results

train network on ITS dataset

python main.py --net='ffa' --crop --crop_size=240 --blocks=19 --gps=3 --bs=2 --lr=0.0001 --trainset='its_train' --testset='its_test' --steps=500000 --eval_step=5000

train network on OTS dataset

python main.py --net='ffa' --crop --crop_size=240 --blocks=19 --gps=3 --bs=2 --lr=0.0001 --trainset='ots_train' --testset='ots_test' --steps=1000000 --eval_step=5000

Test

Trained_models are available at baidudrive: https://pan.baidu.com/s/1-pgSXN6-NXLzmTp21L_qIg with code: 4gat

or google drive: https://drive.google.com/drive/folders/19_lSUPrpLDZl9AyewhHBsHidZEpTMIV5?usp=sharing Put models in the net/trained_models/folder.

Put your images in net/test_imgs/

python test.py --task='its or ots' --test_imgs='test_imgs'

Samples