PyTorch Implementation of "Densely Connected Hierarchical Network for Image Denoising"
Second place winner of sRGB track and Third place winner of Raw-RGB track on NTIRE 2019 Challenge on Real Image Denoising (result paper)
If you find our project useful in your research, please consider citing:
@inproceedings{park2019densely,
title={Densely connected hierarchical network for image denoising},
author={Park, Bumjun and Yu, Songhyun and Jeong, Jechang},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
year={2019}
PyTorch Implementation of "Deep Iterative Down-Up CNN for Image Denoising", second place winner of Raw-RGB track and third place winner of sRGB track is available in here
Python 3.6
PyTorch 1.0.0
MATLAB
We used DIV2K dataset for training. (download)
To generate training patches, use the matlab codes in ./data
We used Kodak and BSD datasets for test datset.
These datasets also need to be modified by matlab codes in ./data to use our test codes.
As an example, use the following command to use our training codes
python main_color.py --batchSize 16 --lr 1e-4 --step 3 --cuda True --train ./data/train.h5 --valid ./data/valid.h5 --gpu 0,1 --checkpoint ./checkpoint
There are other options you can choose.
Please refer to the code.
As an example, use the following command to use our test codes
python test_color.py --cuda True --model1 ./trained.pth --data ./data/noisy.mat --gt ./data/gt.mat --gpu 0 --result ./result/
There are other options you can choose.
Please refer to the code.
To use our pretrained model for gaussian noise, please download here
To use our pretrained model for challenge which is trained with SIDD, please download here
Test results are also available in ./data/results
We retrained our network as we found some problems of our paper version trained parameters.
So the result of the provided pretrained models is a bit different from the paper.
These results can be obtained directly from our test codes.
Color:
Gray:
Note that some results can be a little bit different if you calculate PSNR and SSIM with MATLAB.
If you have any question about the code or paper, please feel free to contact kkbbbj@gmail.com