C3Net

This is a PyTorch implementation of the New Trends in Image Restoration and Enhancement workshop and challenges on image and video restoration and enhancement (NTIRE 2020 with CVPR 2020) paper, C3Net: Demoireing Network Attentive in Channel, Color and Concatenation.

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

@InProceedings{Kim_2020_CVPR_Workshops,
author = {Kim, Sangmin and Nam, Hyungjoon and Kim, Jisu and Jeong, Jechang},
title = {C3Net: Demoireing Network Attentive in Channel, Color and Concatenation},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}

Dependencies

Python 3.6.9
PyTorch 1.4.0

Data

Reference

You have to sign in Codalab and apply to NTIRE 2020 Demoireing Challenge before getting the data.

Proposed algorithm

Training

Use the following command to use our training codes

python train.py

There are other options you can choose. Please refer to train.py.

Test

Use the following command to use our test codes

python test.py

There are other options you can choose. Please refer to test.py.

Performance (PSNR/SSIM)

To use heavier model, we also used numpy to read input data, not hdf5. Hyung-Joon and jisukim helped it.
Our best records can be derived in the code by changing h5 into numpy and reducing GPU memory.

Validation Server	PSNR	SSIM	Rank
Track 1: Single Image	41.30	0.99	9th
Track 2: Burst	40.55	0.99	5th

Testing Server Reference

Testing Server	PSNR	SSIM	Rank
Track 1: Single Image	41.11	0.99	4th
Track 2: Burst	40.33	0.99	5th

Contact

If you have any question about Demoireing model and the CVPR2020 challenge paper, feel free to ask me to [email protected].
If you have any question about Deblurring model, visit here and feel free to ask Hyung-Joon to [email protected].
If you have any question about using more heavier C3Net, visit here and feel free to ask jisukim to [email protected].

Acknowledgement

Thanks for SaoYan who gave the implementaion of DnCNN.
Thanks for yun_yang who gave the implementation of DRRN.
Thanks for BumjunPark who gave the implementation of DHDN.

Hint of color loss from Jorge Pessoa.
Hint of concatenation and residual learning from RDN (informal implementation).
Hint of U-net block from DIDN (formal implementation).

C3Net started from RUN.

More Details

Also, we won 3rd Place in NTIRE 2020 Challenge on Image and Video Deblurring thanks to Hyung-Joon and jisukim.
The code is available at here.