WaterDiff: Mamba-Based Latent Diffusion Model for Underwater Image Enhancement

Abstract

Underwater imaging is affected by absorption and scattering, leading to issues, such as low contrast, poor clarity, color deviation, and haze. To address these problems, we propose a Mamba-based latent diffusion model (LDM) for underwater image enhancement, named WaterDiff, which leverages the potential of LDM and the guidance provided by the transmission map. To alleviate the substantial computational cost incurred by pixel-based diffusion recovery, a latent-based method is utilized, whereby the degraded underwater image is first transformed by a feature encoder and then restored in the compact latent space. To address these challenges, two key components are proposed: first, a feature encoder that extracts the multiscale features enhanced in the decoder to reduce distortions; second, Mamba encoder with linear complexity and transmission map guidance module integrated into the noise predictor to extract structural information, which enables more powerful latent-based image enhancement without significantly increasing inference costs. The experimental results demonstrate the superiority of the proposed WaterDiff and effectiveness of its modules.