Crafting Degradation-Aware Task Affinity Graph for All-in-One Image Restoration
IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17274-17286
DOI: 10.1109/tcsvt.2026.3722683 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Yuxuan Zheng und 6 weiteren Autor:innen. Die Arbeit nennt 80 Literaturangaben. Gefördert von: Joint Funds of the National Natural Science Foundation of China, National Natural Science Foundation of China, Anhui Provincial Natural Science Foundation.
Abstract (Originalsprache)
All-in-one image restoration (AiOIR) has shown great potential in addressing diverse degradations within a unified framework. However, degradation types often vary in semantic compatibility, where some share similar restoration patterns, others stem from fundamentally distinct physical processes, posing inconsistent optimization dynamics, which are frequently overlooked in existing methods. The intrinsic discrepancies between different tasks can result in suboptimal synergy or even interference, ultimately limiting performance on individual degradations. In this work, we propose PalRestorer, which explicitly explores task affinities to harmonize task-shared knowledge and reconcile optimization inconsistencies for AiOIR. Our approach first learns the statistical dependencies between low-quality observations and different degradation knowledge, deriving accurate and deterministic task-specific prompts. It then quantifies inter-task relationships and constructs a degradation-aware affinity graph. This allows aggregating semantically compatible shared knowledge from related tasks while eliminating negative transfer between dissimilar ones, yielding task-sharing pal prompts. Finally, we devise a dynamic modulation module that integrates optimally relevant task knowledge with such pal prompts, establishing a pal-task-informed restoration process. Experiments on diverse specific and composite degradations validate the superiority of our method. The code is available at: https://github.com/zhengyuxuan614/PalRestorer
Förderung
- Joint Funds of the National Natural Science Foundation of China
- National Natural Science Foundation of China
- Anhui Provincial Natural Science Foundation
- Shandong Provincial Natural Science Foundation
- Chinese Institute of Electronics-Tianyi Cloud Research Program
- Intelligent Interconnected Systems Laboratory of Anhui Province
- Fundamental Research Funds for the Central Universities
Zitieren
APA
Zheng, Y., Liu, M., Li, F., Wei, Y., Cong, R., Zhao, Y., & Bai, H. (2026). Crafting Degradation-Aware Task Affinity Graph for All-in-One Image Restoration. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17274-17286. https://doi.org/10.1109/tcsvt.2026.3722683
Vancouver
Zheng Y, Liu M, Li F, Wei Y, Cong R, Zhao Y, et al. Crafting Degradation-Aware Task Affinity Graph for All-in-One Image Restoration. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17274-17286. doi:10.1109/tcsvt.2026.3722683
BibTeX
@article{Zheng2026,
author = {Zheng, Yuxuan and Liu, Man and Li, Feng and Wei, Yanyan and Cong, Runmin and Zhao, Yao and Bai, Huihui},
title = {Crafting Degradation-Aware Task Affinity Graph for All-in-One Image Restoration},
journal = {IEEE Transactions on Circuits and Systems for Video Technology},
year = {2026},
volume = {36},
number = {11},
pages = {17274--17286},
doi = {10.1109/tcsvt.2026.3722683}
}
EUVIMED