Crafting Training Degradation Distribution for the Accuracy-Generalization Trade-off in Real-World Super-Resolution

Ruofan Zhang, Jinjin Gu, Haoyu Chen, Chao Dong, Yulun Zhang, Wenming Yang
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:41078-41091, 2023.

Abstract

Super-resolution (SR) techniques designed for real-world applications commonly encounter two primary challenges: generalization performance and restoration accuracy. We demonstrate that when methods are trained using complex, large-range degradations to enhance generalization, a decline in accuracy is inevitable. However, since the degradation in a certain real-world applications typically exhibits a limited variation range, it becomes feasible to strike a trade-off between generalization performance and testing accuracy within this scope. In this work, we introduce a novel approach to craft training degradation distributions using a small set of reference images. Our strategy is founded upon the binned representation of the degradation space and the Frechet distance between degradation distributions. Our results indicate that the proposed technique significantly improves the performance of test images while preserving generalization capabilities in real-world applications.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-zhang23l, title = {Crafting Training Degradation Distribution for the Accuracy-Generalization Trade-off in Real-World Super-Resolution}, author = {Zhang, Ruofan and Gu, Jinjin and Chen, Haoyu and Dong, Chao and Zhang, Yulun and Yang, Wenming}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {41078--41091}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/zhang23l/zhang23l.pdf}, url = {https://proceedings.mlr.press/v202/zhang23l.html}, abstract = {Super-resolution (SR) techniques designed for real-world applications commonly encounter two primary challenges: generalization performance and restoration accuracy. We demonstrate that when methods are trained using complex, large-range degradations to enhance generalization, a decline in accuracy is inevitable. However, since the degradation in a certain real-world applications typically exhibits a limited variation range, it becomes feasible to strike a trade-off between generalization performance and testing accuracy within this scope. In this work, we introduce a novel approach to craft training degradation distributions using a small set of reference images. Our strategy is founded upon the binned representation of the degradation space and the Frechet distance between degradation distributions. Our results indicate that the proposed technique significantly improves the performance of test images while preserving generalization capabilities in real-world applications.} }
Endnote
%0 Conference Paper %T Crafting Training Degradation Distribution for the Accuracy-Generalization Trade-off in Real-World Super-Resolution %A Ruofan Zhang %A Jinjin Gu %A Haoyu Chen %A Chao Dong %A Yulun Zhang %A Wenming Yang %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-zhang23l %I PMLR %P 41078--41091 %U https://proceedings.mlr.press/v202/zhang23l.html %V 202 %X Super-resolution (SR) techniques designed for real-world applications commonly encounter two primary challenges: generalization performance and restoration accuracy. We demonstrate that when methods are trained using complex, large-range degradations to enhance generalization, a decline in accuracy is inevitable. However, since the degradation in a certain real-world applications typically exhibits a limited variation range, it becomes feasible to strike a trade-off between generalization performance and testing accuracy within this scope. In this work, we introduce a novel approach to craft training degradation distributions using a small set of reference images. Our strategy is founded upon the binned representation of the degradation space and the Frechet distance between degradation distributions. Our results indicate that the proposed technique significantly improves the performance of test images while preserving generalization capabilities in real-world applications.
APA
Zhang, R., Gu, J., Chen, H., Dong, C., Zhang, Y. & Yang, W.. (2023). Crafting Training Degradation Distribution for the Accuracy-Generalization Trade-off in Real-World Super-Resolution. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:41078-41091 Available from https://proceedings.mlr.press/v202/zhang23l.html.

Related Material