DepthAnything and SAM for UIE: Exploring Large Model Information Contributes to Underwater Image Restoration

Abstract Underwater image enhancement (UIE) poses significant challenges due to complex light absorption and scattering effects in aquatic environments, which degrade visual quality and hinder subsequent image analysis. To address these challenges, we propose a novel Large Model-Assisted Feature Enhancement Strategy for UIE that integrates depth and semantic features extracted from large pre-trained models—specifically, DepthAnything and the Segment Anything Model. We develop a rigorous theoretical framework grounded in information theory and statistical learning theory, demonstrating that incorporating auxiliary features from large models enhances the representational capacity of UIE models, reduces estimation risk, and facilitates optimization in high-dimensional feature spaces. Our analysis, leveraging PAC-Bayesian theory, provides theoretical generalization bounds that justify the improved performance of our approach. Extensive experiments on benchmark datasets corroborate our theoretical findings, achieving significant improvements over state-of-the-art methods in both full-reference and no-reference image quality metrics. Through comprehensive ablation studies, we elucidate the critical role of depth information in UIE, aligning with theoretical principles of underwater light attenuation and scattering. This work offers valuable academic insights into integrating large pre-trained models for UIE, bridging the gap between theoretical advancements and practical applications, and opening new avenues for addressing complex image restoration problems in challenging environments.