<p>Underwater image enhancement (UIE) remains a fundamental yet challenging problem in computer vision due to the complex physics of light propagation in aquatic environments. Traditional physics-based or learning-driven approaches often need more prior knowledge and representational capacity to generalize across diverse underwater conditions. This paper presents a novel theoretical framework for leveraging large-scale pre-trained models in UIE, explicitly addressing the fundamental limitations in existing methods through principled integration of depth and semantic priors. Our key contribution is twofold: First, we establish a rigorous information-theoretic foundation that quantifies how auxiliary features from foundation models enhance the representational capacity of UIE systems, providing theoretical guarantees through PAC-Bayesian bounds on generalization performance. Second, we propose a Feature Enhancement Strategy that optimally combines depth information from DepthAnything and semantic priors from the Segment Anything Model, guided by underwater optical physics. We introduce CAB-USRI, a physics-based algorithm for both baseline and theoretical validation. Our extensive experimentation on multiple benchmark datasets demonstrates that our approach consistently outperforms state-of-the-art methods by significant margins while maintaining theoretical interpretability. Our ablation studies reveal the crucial role of depth priors in underwater scenarios, establishing a clear connection between theoretical bounds and empirical performance. This work bridges the gap between foundation models and domain-specific tasks, providing theoretical insights and practical solutions for complex image restoration problems in challenging environments.</p>