Unsupervised Single Image Super-Resolution Network (USISResNet) for Real-World Data Using Generative Adversarial Network

This is repository of code for NTIRE-2020 (CVPRW-2020) paper titled "Unsupervised Single Image Super-Resolution Network (USISResNet) for Real-World Data Using Generative Adversarial Network"

- Framework description

- Results

Method	PSNR(dB)	SSIM	LPIPS
MsDNN	25.08	0.7079	0.482
RCAN	25.31	0.6402	0.576
ESRGAN	19.04	0.2422	0.755
Proposed	21.71	0.5895	0.375

- Test the model To test/reproduce results, change options/test/test_ntire1.json file in which you need to change path for dataset and pre-trained model of G network. Then you need run following command.

python test.py -opt PATH-to-json-file

- Pre-trained model

• The pre-train model is shared in main folder named 116000_G.pth for USISResNet.
• The pre-trained model for QA assessment network trained on KADID dataset as mentioned in the manuscript has also be included as latest_G.pth.

- Required Packages The list of all required packages are included in usisresnet.yml file. You can simply import the .yml file using conda environment.

We are thankful to Xinntao for their ESRGAN code on which we have made this work.

For any problem or query, you may contact to Kalpesh Prajapati at [email protected]

---

Citation

@INPROCEEDINGS{9151093,
  author={K. {Prajapati} and V. {Chudasama} and H. {Patel} and K. {Upla} and R. {Ramachandra} and K. {Raja} and C. {Busch}},
  booktitle={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, 
  title={Unsupervised Single Image Super-Resolution Network (USISResNet) for Real-World Data Using Generative Adversarial Network}, 
  year={2020},
  volume={},
  number={},
  pages={1904-1913},
  doi={10.1109/CVPRW50498.2020.00240}}