Abstract The combination of light absorption and scattering by underwater impurities and inhomogeneous illumination leads to the complex and diverse degradation of underwater environmental images, which inevitably affects the effectiveness of underwater visualization tasks. In recent years, underwater image enhancement techniques, especially neural networkbased methods, have made great progress. However, since acquiring high-quality ground truth images is a major challenge, existing methods can only generate approximate reference maps, thus limiting the processing performance of their networks. Also, it is difficult to create global pixel-to-pixel dependencies since augmented networks often face limited perceptual regions. To address these issues, we propose a novel probability distribution framework, Mamba-SAU, which decomposes UIE into two phases: augmented distribution estimation and consensus refinement, effectively mitigating the approximation error induced by the reference. Drawing inspiration from the selective structured state space model (SSM), which can capture global dependencies with less complexity, we integrate the Mamba module into the underwater image enhancement network to establish global context awareness while maintaining computational efficiency. Experimental results show that our proposed Mamba-SAU approach outperforms state-of-the-art underwater image contrast enhancement algorithms and is validated by subjective and objective evaluation metrics.
Kou et al. (Mon,) studied this question.