Awesome

Scale-arbitrary Image Super-resolution Task

Introduction

After surveying the literature on scale-arbitrary super-resolution, we summarize existing methods into two categories: one-stage and two-stage. The two-stage method involves learning the discrete representation (DR) of an image, requiring the prediction of convolution kernels, while the one-stage method involves learning the continuous representation (CR) of an image, directly mapping coordinates to pixel values.

This study introduces a novel image representation method, named CDCR, which integrates the DR and CR of an image. CDCR is designed as a plug-in module to enable existing super-resolution frameworks to handle arbitrary scale factors. By integrating the advantages of DR and CR, CDCR enhances the overall image accuracy. The CR-based dense prediction module provides a strong initial estimation of the high-resolution image, while the DR-based resolution-specific refinement module fine-tunes the predicted values of local pixels to achieve high accuracy.

We apply our method to various remote sensing datasets, such as RSC11, RSSCN7, and WHU-RS19, achieving advanced performance in this field. Our approach is based on the work of LIIF, but with modifications specific to remote sensing scenarios.

Overall, the proposed CDCR method advances the development of more advanced scale-arbitrary image super-resolution techniques.

Training commands

Train the model directly by running train.py.

python train.py --config configs/train/train_edsr-baseline-cdcr.yaml --gpu 0
python train.py --config configs/train/train_rdn-cdcr.yaml --gpu 0
python train.py --config configs/train/train_rcan-cdcr.yaml --gpu 0

Testing commands

1. The CDCR inference command:

    bash scripts/test-RSC11.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
    bash scripts/test-RSSCN7.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
    bash scripts/test-WHU-RS19.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
   

2. The CDCR+ inference command:

    bash scripts/testSCA-RSC11.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
    bash scripts/testSCA-RSSCN7.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
    bash scripts/testSCA-WHU-RS19.sh save/model_name/epoch_num.pth gpu_num ABCDEFGHIJK
   

   Note that A~K represent scale factors of 2, 3, 4, 6, 8, 12, 16, 20, 3.4, 9.7, and 17.6, respectively.

References

1. LIIF
2. Meta-SR
3. A-LIIF
4. ArbSR

If it helps for you, please cite

@article{an2023combining,
	author={An, Tai and Huo, Chunlei and Xiang, Shiming and Pan, Chunhong},
	title={Combining Discrete and Continuous Representation: Scale-Arbitrary Super-Resolution for Satellite Images},
	journal={Remote Sensing},
	volume={15},
	year={2023},
	number={7},
	article-number={1827},
	url={https://www.mdpi.com/2072-4292/15/7/1827},
	ISSN={2072-4292},
	DOI={10.3390/rs15071827}
}