This repository is the official PyTorch implementation of Better “CMOS” Produces Clearer Images: Learning Space-Variant Blur Estimation for Blind Image Super-Resolution (arxiv). Please feel free to contact us if you have any questions.

• Prepare experimental environments

  pip install -r requirements.txt
  

Note: this repository is based on BasicSR and MANet. Please refer to their repository for a better understanding of the code framework.

To prepare data, download datasets NYUv2-BSR and Cityscapes-BSR from Google Drive and place them in the './datasets' folder. The creation of these two datasets is based on this repository. Please refer to our paper for more detais.

Step1: to train CMOS with NYUv2-BSR, run this command:

python train.py --opt options/train/NYUv2_BSR/train_stage1.yml

Step2: to train non-blind RRDB-SFT with NYUv2-BSR, run this command:

python train.py --opt options/train/NYUv2_BSR/train_stage2.yml

Step3: to fine-tune RRDB-SFT with CMOS, run this command:

python train.py --opt options/train/NYUv2_BSR/train_stage3.yml

All trained models are provided in ./codes/pretrained_models.

To test CMOS (stage1, blur and semantic estimation) with NYUv2-BSR, run this command:

python test.py --opt options/test/NYUv2_BSR/test_stage1.yml

To test RRDB-SFT (stage2, non-blind SR with ground-truth blur and semantic maps) with NYUv2-BSR, run this command:

python test.py --opt options/test/NYUv2_BSR/test_stage2.yml

To test CMOS+RRDB-SFT (stage3, blind SR), run this command:

python test.py --opt options/test/NYUv2_BSR/test_stage3.yml

@inproceedings{chen2023better,
    title={Better" CMOS" Produces Clearer Images: Learning Space-Variant Blur Estimation for Blind Image Super-Resolution},
    author={Chen, Xuhai and Zhang, Jiangning and Xu, Chao and Wang, Yabiao and Wang, Chengjie and Liu, Yong},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    pages={1651--1661},
    year={2023}
}

The codes are based on BasicSR, MANet and MTI-Net. Thanks for their great works.