A Two-Stage Underwater Image Enhancement Network With Cross-Domain Multiple Priors

Abstract

Underwater images in marine environments often suffer from severe visual degradation due to light absorption and scattering. Existing deep learning-based underwater image enhancement (UIE) methods typically rely on limited underwater data sets to learn nonlinear mappings, which constrains their ability to generalize across diverse degradation patterns. Empirically, we find that structural boundaries and texture features from natural images can serve as informative priors for UIE. Motivated by this, we propose a novel two-stage UIE network with cross-domain multiple priors (CMUIE) to improve underwater image quality. First, to address multisource degradation, we design a pretraining-finetuning pipeline that transfers generic enhancement knowledge via blind degradation modeling. Second, to preserve semantic structures, we construct a structure-aware denoising module guided by semantic priors, directing the network's focus to salient regions. Finally, to restore high-frequency textures, we develop a one-step diffusion reconstruction module that leverages a diffusion prior to enhance fine-grained details. Extensive experiments show that CMUIE achieves state-of-the-art performance in both image denoising and super-resolution tasks, exhibiting superior generalization capabilities and practical applicability.