Include：SRCNN、FSRCNN、SRResNet、SRGAN

There are many Pytorch implementations of these networks on the web, but they do not appear exactly as described in the paper, so the results are quite different from the paper. So I've provided a version that's as close to the paper as possible.
Hopefully it will help those interested in Super-Resolution networks get started.
For more details:https://zhuanlan.zhihu.com/p/431724297

You can clone this repository directly, and run it without installing.

• Pytorch 3.7 64bit
• Windows 10

Because bicubic interpolation in python is different with matlab, but in paper use matlab to generate datasets and evaluate PSNR, so I found a Python implementations of Matlab function:imresize().
Here is the author's repository.
Similiarly, I give a python version of the rgb2ycbcr() and ycbcr2rgb() in matlab.

• Run data_aug.py to augment datasets.
• Run gen_datasets.py to generate trainning and validation data. (You may need to modify parameters in config.)

Take SRCNN as an example, run SRCNN_x2.py to train SRCNN. You can modify the training parameters according to your needs follow this template.

Run test.py to generate test result and calculate PSNR. (You can modify parameters to specify testsets.)

Run csv2visdom.py can visualize converge curve with visdom. (You need to install visdom and run it in advance.)
Then visit localhost:8097.

Train on 91-images.

Train on DIV2K.

Train on DIV2K.

SRGAN cannot be evaluated by PSNR alone, so I list some test result.
Obviously, SRGAN generates a sharper results than SRResNet and looks more convincing.

bicubic	SRResNet	SRGAN	original