FouriDown, as a generic operator, comprises four key components: 2D discrete Fourier transform, context shuffling rules, Fourier weighting-adaptively superposing rules, and 2D inverse Fourier transform. These components can be easily integrated into existing image restoration networks.
The model equipped with FouriDown generates much unique and strong global responses. In contrast, the model with other down-sampling method responds weakly to these regions.
Compared to other methods, our FouriDown adaptively adjusts the high and low frequencies, resulting in a wider-band response in the output feature spectrum. Contrasted with previous methods that used fixed frequency aliasing patterns, our approach activates a broader bandwidth on the spectrum, bringing the enhanced performance in image restoration.
If you have any problems with the released code, please do not hesitate to contact me by email ([email protected]).
If you find this project useful in your research, please consider cite:
@inproceedings{zhu2023fouridown,
title={FouriDown: Factoring Down-Sampling into Shuffling and Superposing},
author={Zhu, Qi and Zhou, Man and Huang, Jie and Zheng, Naishan and Gao, Hongzhi and Li, Chongyi and Xu, Yuan and Zhao, Feng},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023}
}