[ICME 2023] LKD-Net: Large Kernel Convolution Network for Single Image Dehazing

This is the official implementation of LKD-Net.

Network Architecture

Getting Started

Installation

Clone our repository.

git clone https://github.com/SWU-CS-MediaLab/LKD-Net.git
cd LKD-Net

Make conda environment.

conda create -n LKD python=3.7.0
conda activate LKD

Install dependencies.

conda conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=10.2 -c pytorch
pip install -r requirement.txt

Datasets Preparation

RESIDE official website here.

DATASETS FILE STRUCTURE

|-- data
    |-- ITS
        |-- hazy
            |-- *.png
        |-- gt
            |-- *.png
    |-- OTS 
        |-- hazy
            |-- *.jpg
        |-- gt
            |-- *.jpg
    |-- SOTS
        |-- indoor
            |-- hazy
                |-- *.png
            |-- gt
                |-- *.png
        |-- outdoor
            |-- hazy
                |-- *.jpg
            |-- gt
                |-- *.png

Train

Run the following script to train you own model.

python train.py \
--model LKD-t \
--model_name LKD.py \
--num_workers 8 \
--save_dir ./result \
--datasets_dir ./data \
--train_dataset ITS \
--valid_dataset SOTS \
--exp_config indoor \
--gpu 0 \
--exp_name train_my_model \

Test

Run the following script to test the trained mode.

python test.py --model (model name) --model_weight (model weight dir) --data_dir (your dataset dir) --save_dir (path to save test result) --dataset (dataset name) --subset (subset name)

For example, we test the LKD-t on the SOTS indoor set.

python test.py --model LKD-t --model_weight ./result/RESIDE-IN/LKD-t/LKD-t.pth --data_dir ./data --save_dir ./result --dataset SOTS --subset indoor

Acknowledgement

Thanks for the work of Yuda Song et al. Our code is heavily borrowed from the implementation of Dehazeformer.