LHNet: A Low-cost Hybrid Network for Single Image Dehazing

News

• [2024.1.09] We have comprehensively improved the performance of LHNetV1, especially the inference speed, and latest paper is currently under review. The code of LHNetV2 will be released after the latest paper is accepted.

Overview

This repository contains the offical PyTorch implementation of paper:

LHNet: A Low-cost Hybrid Network for Single Image Dehazing, ACM MM 2023

[Paper Link]

Dependences

1. Clone Repo

   git clone https://github.com/SHYuanBest/LHNet-ACM-MM-23.git
   cd LHNet-ACM-MM-23-main/
   

2. Create a virtual environment via conda.

   conda create -n LHNet python=3.7
   conda activate LHNet

3. Install requirements.

   pip install -r requirements.txt

4. Download pre-trained models.

   • Download the pre-trained generator and discriminator from here, and place it into the folder ./ckpts

     ckpts
        |- ITS
           |- ITS.pt
        |- OTS
           |- OTS.pt
        |- Dense
           |- Dense.pt  
        |- NH
           |- NH.pt
     

5. Data preparing.

   • LHNet is trained on RESIDE-ITS, RESIDE-OTS, DENSE-HAZE and NH-HAZE.

   • Download all the datasets and place it into the folder ./datasets.

     Dataset	RESIDE-ITS	RESIDE-OTS	Dense-Haze	NH-HAZE	SOTS-ITS/OTS
     Details	about 4.5G	about 45G	about 250M	about 315M	about 415M
     Google drive	Google drive	Google drive	Google drive	Google drive	Google drive

     The directory structure will be arranged as:

     datasets
        |-ITS
           |- train_ITS
              |- haze
              |- clear_images
              |- trainlist.txt
           |- valid_ITS
              |- input
              |- gt
              |- val_list.txt
        |-OTS   
           |- train_OTS
              |- haze
              |- clear_images
              |- trainlist.txt
           |- valid_OTS
              |- input
              |- gt
              |- val_list.txt
        |-Dense-Haze
           |- train_Dense
              |- haze
              |- clear_images
              |- trainlist.txt
           |- valid_Dense
              |- input 
              |- gt
              |- val_list.txt
        |-NH-Haze
           |- train_NH
              |- haze
              |- clear_images
              |- trainlist.txt
           |- valid_NH
              |- input 
              |- gt
              |- val_list.txt
     

   • Or you can customize the dataset to follow the data format as shown before.

Training

See python src/train.py --h for list of optional arguments.

An example:

CUDA_VISIBLE_DEVICES=0 python src/train.py \
  --dataset-name your_choice \
  --data-dir ../datasets \
  --ckpt-save-path ../weights \
  --ckpt-load-path ../ckpts/your_choice.pt
  --nb-epochs 2000 \
  --batch-size 8 \
  --train-size 512 512 \
  --valid-size 512 512 \
  --loss l1 \
  --plot-stats \
  --cuda True

Evaluation

See python src/test.py --h for list of optional arguments.

An example:

CUDA_VISIBLE_DEVICES=0 python src/test.py \
  --dataset-name your_choice \
  --data-dir your_choice \
  --ckpts-dir your_choice \
  -val_batch_size 1

Results

Citing LHNet

If you find LHNet useful in your research, please consider citing our paper:

@inproceedings{yuan2023lhnet,
  title={LHNet: A Low-cost Hybrid Network for Single Image Dehazing},
  author={Yuan, Shenghai and Chen, Jijia and Li, Jiaqi and Jiang, Wenchao and Guo, Song},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={7706--7717},
  year={2023}
}

Contact

If you have any question, please feel free to contact us via [email protected]

Acknowledge

This code is based on UNet, Swin-Transformer and Dehamer.