RDGAN

A Pytorch Implementation of RDN, and SRGAN+RDN

Code Introduction

We have done some experiments on the super-resolution and simultaneously denoising task. One problem for super-resolution and simultaneously denoising task is that the blur after denoising will be amplified by the super-resolution. What we have tried is to utilize GAN to generate some details of the blured part.
We established two super-resolution architecture, one is RDN, and the other is SRGAN. A small difference is that the generator of SRGAN is RDN, taking advantage of superior effect of super resolution of RDN.
Codes comprise of three parts, the first part is the data generator. 5 small patches are randomly selected from DIV2K training datasets, totally 4000 patches. And whole image is preserved for DIV2K validation dataset. The second part is model. The third part is training and testing code, among which train.py is the training code of SRGAN+RDN, train_rdn.py is the training code of RDN. And test_image.py is the testing code of these models.

Results

Bicubic Downsample x3 RDN

Dataset	PSNR(dB)	SSIM	BRISQUE
Set5	35.0033	0.9467	32.826
Set14	31.1504	0.8822	37.7410

Bicubic Downsample x3 + Gaussian Noise Level 25 RDN

Dataset	PSNR(dB)	SSIM	BRISQUE
Set5	29.3928	0.8418	47.2392
Set14	27.4011	0.7577	46.4806

Bicubic Downsample x3 + Gaussian Noise Level 25 SRGAN+RDN

Dataset	PSNR(dB)	SSIM	BRISQUE
Set5	29.3432	0.8392	47.4282
Set14	27.3637	0.7534	44.2279

PSNR and SSIM are calculated by the Y channel of YCrCb image converted from RGB image.