Content-decoupled Contrastive Learning-based Implicit Degradation Modeling for Blind Image Super-Resolution

[paper]

Created by Jiang Yuan, Ji Ma, Bo Wang, Weiming Hu

This repository contains PyTorch implementation for Content-decoupled Contrastive Learning-based Implicit Degradation Modeling for Blind Image Super-Resolution (Accepted by IEEE TIP 2025).

Train

1. Prepare training data

1.1 Download the DIV2K dataset and the Flickr2K dataset.

1.2 Combine the HR images from these two datasets in ./datasets/DF2K/HR to build the DF2K dataset.

2. change degradation config

Change the TODO section in main.sh, option.py and trainer.py to select the corresponding degradation settings.

3. Begin to train

Run main.sh to train on the DF2K dataset.

Test

1. Prepare test data

Download benchmark datasets (e.g., Set5, Set14 and other test sets) and prepare HR/LR images in ./datasets/benchmark.

2. change degradation config

Change the TODO section in test.sh, option.py and trainer.py to select the corresponding degradation settings.

3. Begin to test

Run test.sh to test on benchmark datasets.

Citation

@article{10964088,
  author={Yuan, Jiang and Ma, Ji and Wang, Bo and Hu, Weiming},
  journal={IEEE Transactions on Image Processing}, 
  title={Content-Decoupled Contrastive Learning-Based Implicit Degradation Modeling for Blind Image Super-Resolution}, 
  year={2025},
  volume={34},
  pages={4751-4766},
  doi={10.1109/TIP.2025.3558442}}