[✨AAAI 2025] DRAT: Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution

Sen Chen, Hongying Liu, Chaowei Fang, Fanhua Shang, Yuanyuan Liu, Liang Wan, Dongmei Jiang, Yaowei Wang

School of Artiffcial Intelligence, Xidian University

News:

• Our DRAT is accepted by AAAI2025✨!
• paper: https://doi.org/10.1609/aaai.v39i2.32216

Overview of DRAT

Environment

This code is based on basicsr.

• python >= 3.9
• pytorch == 1.13.1

conda env create -n DRAT -f environment.yml

python setup.py develop

Code File Descriptions

File	Description
basicsr/archs/drat_arch.py	Implementation of DRAT.
basicsr/add/encoder.py	Implementation of aux and blur encoder.
basicsr/losses/basic_loss.py	Definitions of the loss functions.
degradation/scripts/	Code for generating training and test data for Isotropic, Anisotropic Gaussian kernels and complex degradations.
inference/inference_drat.py	Code for inferring Super-Resolution (SR) results by using DRAT.
options/test/xxx/xxx.yml	YML configuration file for testing.
options/train/xxx/xxx.yml	YML configuration file for training.
basicsr/utils.py	Various utilities.
options/train/xxx/xxx.yml	YML configuration file for training.
moco/	Using MoCo to train Encoder for degradation representation learning
checkpoints/	Pretrained models

You can use the following commands to train and test:

Dataset Prearation

• We use DIV2K and Flickr2K as our training datasets (totally 3450 images).

• For more details about the dataset, please see DatasetPreparation, which has a detailed introduction.

• For Anisotropic Gaussian kernels degradation, you can download DIV2KRK test datasets from DIV2KRK.

• For real-world degradation test, you can download the RealSR_V3(Nikon and Canon), DRealSR and ImageNet256_x4_sr (from ResShift)

cd degradation/scripts/
python extract_subimages.py # crop the images to subimages

# getting degradation training data
python generate_blur_traindata.py # for iso/aniso blind sr (for training)
python generate_realesrgan_data.py # for real-world sr (for training)
python generate_blur_testdata.py # for generate iso/aniso test datasets

Pretrained Model

• You can download pretrained models from Pretrained_DRAT.

Train

• The alorithm is in archs/drat_arch.py.
• Please Modify options/train/SettingX/xx.yml to set path, iterations, and other parameters.
• To train the DRAT, run the commands below.

python train.py -opt=options/train/DRAT/setting1/x4/train_DRAT_x4.yml
python train.py -opt=options/train/DRAT/setting2/x4/train_DRAT_div2krkx2.yml
python train.py -opt=options/train/DRAT/setting3/x4/train_DRATGAN.yml

Test & Infer

• Before test, please modify the model and data path.
• To test the DRAT, run the commands below.

python test.py -opt=options/test/DRAT/setting1/x4/test_urban100.yml
python test.py -opt=options/test/DRAT/setting2/x4/test_div2krkx4.yml
python test.py -opt=options/test/DRAT/setting2/x4/test_DRATGAN.yml

• To infer DRAT on other image folders.
• To infer the DRAT, run the commands below.
• cd inference/

python inference_drat.py --input your_input_path --output your_save_path

Result

Comparison of Isotropic Gaussian Kernels (Gaussian8)

Comparison of Anisotropic Gaussian Kernels (DIV2KRK)

Comparison of Real-World Degradation (RealSR)

Other Super-Resolution Visualization

Citations

If our code helps your research or work, please consider citing our paper. The following is a BibTeX reference.

@inproceedings{hqqh✨,
  title={Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution},
  author={Snd},
  booktitle={这棵🌳成精后，天天求我教它跳广场舞},
  pages={1818-8984},
  year={2025}
}

Contact

If you have any questions, please feel free to reach me out at 23171214481@stu.xidian.edu.cn.

Acknowledgement

Our projects is based on DASR, DCLS and BasicSR. Thanks for their awesome works.