Codes for Semi-supervised single image dehazing.
pytorch >= 1.0visdom
Make you dataset by:
- synthetic images: align two images (hazy(HxWxC), clean(HxWxC)) into one image (Hx2WxC). To be noted that H and W should be multiple of 8. Put them (~6000 images) in
./datasets/dehazing/train. - real hazy images: put them(~1000 images) in
./datasets/dehazing/unlabeled - test images: align the same as 1. and put them in
./datasets/dehazing/test
The you can train the model by:
python train.py --dataroot ./datasets/dehazing --name run_id_1 --learn_residual --display_freq 100 --print_freq 100 --display_port 8097 --which_model_netG EDskipconn --lambda_vgg 10 --lambda_mse 1000 --lambda_ssim 0 --niter 90 --niter_decay 0 --fineSize 256 --no_html --lambda_DC 1e-4 --lambda_TV 1e-4 --gpu_id 0 --update_ratio 1 --unlabel_decay 0.99 --save_epoch_freq 1 --semi --patch_size 7 --batch_size 2
Monitor the traning process via visdom by:
python -m visdom.server -port 8097
You can test you model on RESIDE SOTS dataset.
python test.py --dataset ./datasets/dehazing --name run_id_1 --learn_residual --which_model_netG EDskipconn --gpu_id 0 --no_html --which_epoch latest
@article{li2019semi,
title={Semi-supervised image dehazing},
author={Li, Lerenhan and Dong, Yunlong and Ren, Wenqi and Pan, Jinshan and Gao, Changxin and Sang, Nong and Yang, Ming-Hsuan},
journal={IEEE Transactions on Image Processing},
volume={29},
pages={2766--2779},
year={2019},
publisher={IEEE}
}