Learning Privileged Degradation Priors for All-in-One Image Restoration

Ye Deng, Zhihao Li, Wenli Huang, YangWu, Jiang Duan

Keywords: All-in-One Image Restoration, Privileged Learning

Abstract: The central challenge in all-in-one image restoration lies in learning degradation-specific priors to effectively modulate a restoration network. Prevailing approaches tackle this by learning representations that can distinguish between degradation types, often via proxy tasks like classification or contrastive learning. However, a representation optimized for discrimination is not necessarily optimal for restoration, leading to a fundamental objective mismatch. To address this, we introduce the Learning Using Privileged Information (LUPI) paradigm. Our method employs a teacher network granted privileged access to both degraded and clean images during training, allowing it to learn a prior directly guided by the final restoration quality. This process yields an ideal, inherently ``restoration-aware" prior, which a student network—observing only the degraded input—is then trained to approximate. The learned prior dynamically modulates a restoration backbone for adaptive recovery, enabling our unified model to achieve state-of-the-art performance on benchmarks. Visualizations confirm the learned prior space is semantically structured, revealing intrinsic relationships between degradation types and effectively distinguishing their intensities. The code will be made publicly available upon acceptance of the paper.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 8416