SuperResolution_AMFFN_TGRS2023

"Lightweight Remote-Sensing Image Super-Resolution via Attention-Based Multilevel Feature Fusion Network," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-15, 2023 (https://github.com/cslxju/SuperResolution_AMFFN_TGRS2023)

https://ieeexplore.ieee.org/document/10328614

Requirements

• Python 3.8
• Pytorch=1.12.0
• torchvision=0.13.0
• matplotlib
• opencv-python
• scipy
• tqdm
• scikit-image
• thop

Installation

Clone or download this code and install aforementioned requirements

cd codes

Dataset

We used the UCMerced and RSSCN7 dataset for both training and testing. Download the UCMerced dataset [Google Drive], and RSCNN7 dataset [Google Drive].

These datasets are for training and testing which have been split them into train/val/test data. In every dataset, the original images would be taken as the HR references and the corresponding LR images are generated by bicubic down-sample.

Download the results

We share the trained AMFFN model [Google Drive]. We also share the super-resolved results generated by our AMFFN. Then, researchers can compare their algorithms to our AMFFN without performing inference. Results are available at [Google Drive].

Train

The train/val data pathes are set in data/init.py

# x8
python demo_train.py --model=AMFFN --dataset=UCMerced --scale=8 --image_size=256 --patch_size=192 --ext=img --save=AMFFNx8_UCMerced
# x4
python demo_train.py --model=AMFFN --dataset=UCMerced --scale=4 --image_size=256 --patch_size=192 --ext=img --save=AMFFNx4_UCMerced
# x3
python demo_train.py --model=AMFFN --dataset=UCMerced --scale=3 --image_size=255 --patch_size=144 --ext=img --save=AMFFNx3_UCMerced
# x2
python demo_train.py --model=AMFFN --dataset=UCMerced --scale=2 --image_size=256 --patch_size=96 --ext=img --save=AMFFNx2_UCMerced

Test

The test data path and the save path can be edited in demo_deploy.py

# x8
python demo_deploy.py --model=AMFFN --scale=8
# x4
python demo_deploy.py --model=AMFFN --scale=4
# x3
python demo_deploy.py --model=AMFFN --scale=3
# x2
python demo_deploy.py --model=AMFFN --scale=2

Evaluation

Compute the evaluated results in term of PSNR and SSIM, where the SR/HR paths can be edited in calculate_PSNR_SSIM.py

cd metric_scripts 
python calculate_PSNR_SSIM.py

Citation

If you find this work helpful, please consider citing the following paper:

@ARTICLE{10328614,
  author={Wang, Hongyuan and Cheng, Shuli and Li, Yongming and Du, Anyu},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Lightweight Remote-Sensing Image Super-Resolution via Attention-Based Multilevel Feature Fusion Network}, 
  year={2023},
  volume={61},
  number={1},
  pages={1-15}
}


## Acknowledgements 
This code is built on [HSENet (Pytorch)](https://github.com/Shaosifan/HSENet), [CTNet (Pytorch)](https://github.com/BITszwang/CTNet) and [HAUNet (Pytorch)](https://github.com/likakakaka/HAUNet_RSISR).  We thank the authors for sharing the codes.

## Citation  
[1] H. Wang, S. Cheng*, Y. Li and A. Du, "Lightweight Remote-Sensing Image Super-Resolution via Attention-Based Multilevel Feature Fusion Network," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-15, 2023 (https://github.com/cslxju/SuperResolution_AMFFN_TGRS2023)