Skip to content

FFA-Net: Feature Fusion Attention Network for Single Image Dehazing

Notifications You must be signed in to change notification settings

zhilin007/FFA-Net

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

41 Commits
 
 
 
 
 
 
 
 

Repository files navigation

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

About

FFA-Net: Feature Fusion Attention Network for Single Image Dehazing

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages