Underwater Image Enhancement Combining State-Space Models and Physics-Guided Approaches

Underwater images suffer from degradation issues such as low contrast and color deviation due to light absorption and scattering, making it challenging for existing methods to simultaneously account for physical imaging mechanisms and global feature capture. To address this, this paper proposes an enhancement network that integrates state space models with depth-aware physical priors. The method leverages a visual state space architecture with linear complexity to efficiently extract global features, estimates physical parameters such as transmission by incorporating underwater spectral attenuation patterns, and employs a multi-stage physically guided cascaded unfolding refinement framework to progressively correct color deviations and texture details. Additionally, a semantic-aware loss is introduced to enhance the visual naturalness of the results. Experimental findings demonstrate that this method exhibits outstanding performance across six datasets, including UIEB and LSUI. Specifically, on the LSUI dataset, it achieves a PSNR as high as 28.40 dB, outperforming existing mainstream algorithms in dehazing and color restoration while incurring lower computational costs. This work presents an efficient and physically interpretable solution for underwater image enhancement.