FourierSR: A Fourier Token-based Plugin for Efficient Image Super-Resolution

Paper (ArXiv) | Paper (IEEE)

---

Contents

The contents of this repository are as follows:

1. Dependencies
2. Train
3. Test

Dataset

We used only the first 800 images of DIV2K dataset to train our model.

The test set including Set5, Set14, B100, Urban100, Manga109, which can be downloaded from here.[Password:8888]

The code and datasets need satisfy the following structures:

├── DMNet  					# Train / Test Code
├── dataset  					# all datasets for this code
|  └── DIV2K_decoded  		#  train datasets with npy format
|  |  └── DIV2K_train_HR  		
|  |  └── DIV2K_train_LR_bicubic 			
|  └── benchmark  		#  test datasets with png format 
|  |  └── Set5
|  |  └── Set14
|  |  └── B100
|  |  └── Urban100
|  |  └── Manga109
 ─────────────────

---

Results

Our SR Results can be downloaded from here.

Pretrained models can be found in experiments.

---

Dependencies

• Python >= 3.7
• torch >= 1.2
• einops
• timm
• tqdm
• imageio

Train

# For X2
python3 main.py --model edsr_fre --scale 2 --patch_size 96 --extra_loss --save Offical_EDSRfrex2

# For X3
python3 main.py --model edsr_fre --scale 3 --patch_size 144 --extra_loss --save Offical_EDSRfrex3

# For X4
python3 main.py --model edsr_fre --scale 4 --patch_size 192 --extra_loss --save Offical_EDSRfrex4

Test

# For X2
python3 main.py --model edsr_fre --scale 2 --data_test Set5+Set14+B100+Urban100+Manga109 --save_results --save test_results/EDSRfrex2_results --pre_train ../experiment/Offical_EDSRfrex2/edsr_fre_x2.pt

# For X3
python3 main.py --model edsr_fre --scale 3 --data_test Set5+Set14+B100+Urban100+Manga109 --save_results --save test_results/EDSRfrex2_results --pre_train ../experiment/Offical_EDSRfrex3/edsr_fre_x3.pt

# For X4
python3 main.py --model edsr_fre --scale 4 --data_test Set5+Set14+B100+Urban100+Manga109 --save_results --save test_results/EDSRfrex2_results --pre_train ../experiment/Offical_EDSRfrex4/edsr_fre_x4.pt

Citation

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{li2026fouriersr,
  title={FourierSR: A Fourier Token-based Plugin for Efficient Image Super-Resolution},
  author={Li, Wenjie and Guo, Heng and Hou, Yuefeng and Ma, Zhanyu},
  journal={IEEE Transactions on Image Processing},
  year={2026}
}

Acknowledgement

The foundation for the training process is profited from the outstanding contribution of EDSR.

Contact

This repo is currently maintained by [email protected] and is for academic research use only.