摘要
Recent localization-guided restoration methods have demonstrated promising performance by providing explicit degradation-aware restoration guidance for adverse-weather image restoration. However, reliable degradation localization remains challenging under heterogeneous weather conditions. Although existing studies primarily focus on improving localization through network design and training strategies, whether localization reliability is also constrained by representation quality is less well understood. To address this issue, we propose REAL, an unpaired all-in-one bad-weather restoration framework that revisits degradation localization from a representation perspective. Since degradation masks are estimated directly from image representations, localization reliability inherently depends on representation quality. Therefore, REAL introduces a Weather-guided Representation Adaptation Module (WRAM) to perform representation adaptation before localization, enabling more reliable degradation localization under heterogeneous weather conditions. Furthermore, we propose a Localization Trajectory Learning (LTL) strategy that transfers the benefits of iterative localization refinement during training into efficient single-step localization during inference. Experiments on synthetic and real-world weather datasets demonstrate that REAL achieves competitive restoration performance. Compared with degraded inputs, REAL improves average object detection and semantic segmentation performance by 24% and 32%, respectively, highlighting its practical benefit for autonomous driving perception.
•Develops an unpaired framework for all-in-one adverse-weather removal.•Revisits degradation localization through the lens of representation quality.•Representation adaptation boosts localization reliability under mixed degradations.•Localization trajectory learning progressively refines degradation awareness.•Improved restoration enhances perception robustness for autonomous driving.