Dowload data set from the link: : https://www.dropbox.com/s/wc3b0q0d3querb3/Dehazing_datasets.zip?dl=0
Create data folder:
mkdir data
Unzip dataset to data folder such that we have:
- data/IndoorTestHazy
- data/IndoorTrainGT
- data/IndoorTrainHazy
- data/OutdoorTestHazy
- data/OutdoorTrainGT
- data/OutdoorTrainHazy
Set up environment:
conda create -n dehaze python=3.6
conda activate dehaze
pip install -r requirement.txt
Train the network by run corresponding command below:
Indoor:
./net_train_indoor.sh
Outdoor:
./net_train_outdoor.sh
I provide pretrained model at url: https://drive.google.com/file/d/1WfsmkGmo504ZI7V19_t-euKDNqwW0woC/view?usp=sharing
upzip the pretrained model to src folder such that we have these folders:
- resultIn/Net1/model/model_best.pt
- resultOut/Net1/model/model_best.pt
Run test script to generate output images:
./net_test_in_out.sh
All the result will be store in val folder
In case that you want to test your model, read the test_model.sh and modify the pretrained_model path.
You can download MATLAB evaluation code at this link: https://www.dropbox.com/s/xpcqcucxjn2y28d/evaluation_code.zip?dl=0
Copy your output images into Input folder and run matlab file: evaluate_results.m to get NIQE score
| Indoor (NIQE) | Outdoor(NIQE) | |
|---|---|---|
| HAZY | 6.4564 | 4.1471 |
| OUR | 3.6753 | 3.6608 |
Statistic on 1 GPU Titan X
| Indoor | Outdoor | |
|---|---|---|
| Generator parameter | 34.1M | 34M |
| Discriminator parameter | 5.5M | 5.5M |
| Training time (10000 epoches) | 52.9 hour | 61.0 hour |
| Testing time | 0.0241 | 0.1765 |