Implement-USRNet

This repository is an unofficial implementation of the paper Deep Unfolding Network for Image Super-Resolution

File Structure

File/Folder	Description
img/	img for readme
options/	configuration files
kernels/	pre-trained kernels
model_zoo/	pre-trained models
models/	model files
testsets/	test images
utils/	utility functions
main_test_bicubic.py main_test_realapplication.py main_test_table1.py	test code for images and tables in the paper
USRNet.yaml	conda environment file
train.py	training code
test.py	test code

paper translation

Instructions

1. Clone the repository
2. Create a conda environment using the provided USRNet.yaml file
3. Prepare the training and testing data:
   • For training, use DIV2K+Flickr2K datasets
   • For testing, use 100 images from ImageNet (you can download from here)
4. Prepare the pre-trained models:
   • you can download pretrained models from official pretrained models or my pretrained models
   • put the pretrained models in model_zoo/ folder
5. Change the paths in options/train_usrnet.json:
   • "datasets"/"train"/"dataroot_H": your train data folder
   • "datasets"/"test"/"dataroot_H": your test data folder
   • "path"/"root"/"pretrained_netG": your pre-trained model path（'xxx/Implement-USRNet/model_zoo/xxx.pth'）

Reminder

I use wandb to log the training process. You can change wandbconfig to True in test.py and train.py if you want to use it.

When run main_test_bicubic.py main_test_realapplication.py main_test_table1.py, you need to change the model_name variable in the code to the name of the pre-trained model in the model_zoo folder.

Visualize training Loss/PSNR

Citation

@inproceedings{zhang2020deep, % USRNet
  title={Deep unfolding network for image super-resolution},
  author={Zhang, Kai and Van Gool, Luc and Timofte, Radu},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  pages={3217--3226},
  year={2020}
}