DARVIS: Degradation-Aware Vision-Language Guidance for Adaptive Underwater Image Restoration
Underwater images often suffer from multiple coexisting degradations, including blur, noise, color casts, low contrast, and scattering-induced visibility loss. Since these degradations vary significantly across scenes, a fixed restoration strategy may not always respond effectively to different underwater conditions. This chapter presents Degradation-Aware Restoration with Vision-Language Guidance (DARVIS), a framework that explores how vision-language models can be used to characterize underwater degradations and guide adaptive image restoration. DARVIS employs the Bootstrapped Language-Image Pretraining-2 (BLIP-2) model to generate semantic image descriptions, which are combined with degradation-specific cues to form degradation aware captions. These captions are used to adapt a Contrastive Language-Image Pretraining (CLIP) based degradation classifier for estimating artifact types and severities, such as blur, noise, contrast loss, and color shifts. The predicted degradation profile is then mapped to degradation-aware weights for guiding expert based restoration within a Mixture-of-Experts (MoE) module. The focus of this chapter is on degradation-aware vision-language guidance and expert-based restoration routing rather than on a single fixed restoration backbone. A lightweight restoration backbone is used as one possible implementation within the proposed adaptive framework. Experimental analysis on standard underwater image enhancement benchmarks is used to illustrate then potential of degradation-aware routing for improving restoration adaptability under varying underwater imaging conditions