Degradation-Guided Underwater Image Restoration with Task-Oriented Latent Control
This work addresses the limitation of existing underwater image restoration methods in effectively leveraging the dual role of degradation cues—as both guidance and interference. To this end, we propose the PROTEUS framework, which integrates degradation-guided dynamic feature modulation with a task-oriented latent control mechanism. Our approach enables spatially adaptive feature modulation at the feature level and introduces structured control codes to regulate skip connections at the channel level in the representation space. Notably, the method operates without requiring clean reference embeddings; instead, it learns control codes through discriminative regularization, facilitating multi-stage adaptive restoration. Extensive experiments demonstrate state-of-the-art performance across five paired and four no-reference benchmarks, achieving an effective balance between restoration quality and computational efficiency.