Physical-aware uncertainty prompt learning for real-world blind image restoration

Abstract

Recent years have witnessed notable progress in universal image restoration, which tackles multiple image degradations using a single model. However, these methods struggle to handle complex real-world scenarios due to the lack of paired real data and limited adaptability to unknown corruptions. To address these challenges, we propose a novel Physical-aware Uncertainty Prompt (PUP) paradigm for real-world blind image restoration. Specifically, instead of simply employing pre-synthesized degraded images, we develop a Physical-aware Degradation Modeling scheme (PDM) that considers multiple distortion factors to generate more authentic degraded data online during training. To adaptively handle unknown corruptions, we propose an Uncertainty-Prompted Fourier Transformer (UPFomer) for unified image restoration, which comprises two collaborative designs: Spatial-Frequency Selective Interaction (SSI) and Uncertainty Prompt Alignment (UPA). The former aggregates global frequency information and local spatial context for robust feature representation, while the latter interacts learnable prompts with SSI features via uncertainty weights to compute degradation-aware knowledge. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods on real-world images while ensuring favorable model efficiency.