DCLRNet: Dynamic-Coupled Lightweight Restoration Network for Improving Underwater Visual Quality

Abstract

Water absorption, scattering, and nonuniform illumination jointly cause severe color shifts, visibility loss, and texture degradation in underwater images, while different degradation factors often exhibit complex spatial coupling. Existing methods mainly focus on specific degradation types or single-scale restoration, making it difficult to simultaneously achieve high-quality reconstruction, natural visual appearance, and efficient inference under mixed degradations and limited computational resources. To address these challenges, we propose a dynamic-coupled lightweight restoration network (DCLRNet) for improving underwater visual quality in complex environments. Specifically, the network first constructs an adaptive hybrid representation from the input image and combines it with transmission priors to generate multistage prompts, which dynamically inject degradation cues into feature restoration at different scales. Subsequently, we introduce a prompt-modulated wavelet restoration unit that establishes lightweight coupling between spatial structures and frequency-domain details to jointly recover local textures and high-frequency information. Meanwhile, we develop a dynamic degradation-component restoration mechanism for mixed and spatially varying degradations, which uses degradation characterization and severity-aware routing to adaptively select restoration branches and adjust their response strengths according to the input conditions. Finally, multistage residual information is progressively coupled to aggregate effective restoration cues and generate visually improved restoration results with natural colors and enhanced structural fidelity. Extensive experiments on multiple public underwater image enhancement datasets demonstrate that DCLRNet achieves stable and competitive performance while providing a favorable balance among restoration quality, visual fidelity, and computational efficiency.