Region Gradient-Guided Diffusion Model for Underwater Image Enhancement

Underwater image enhancement (UIE) is a critical challenge in marine visual perception and underwater robotics due to complex aquatic environments that severely degrade image quality. This paper introduces the Region Gradient-Guided Diffusion Model (RGGDM), a novel framework that addresses the fundamental trade-off between local detail preservation and global consistency in UIE. RGGDM innovatively integrates a region gradient-guided mechanism with a hybrid Swin-ConvNeXt architecture, introducing a spatially adaptive denoising process governed by gradient discrepancies between input and target images. We propose a learnable parameter $\delta$ that dynamically modulates denoising intensity, focusing computational resources on semantically salient regions. Our approach is underpinned by rigorous mathematical analysis, demonstrating convergence properties under mild assumptions and providing theoretical guarantees for the model's stability and effectiveness. The synergistic combination of Swin Transformer and ConvNeXt enhances feature representation, significantly improving both perceptual quality and pixel-level accuracy. Extensive experiments on benchmark datasets demonstrate RGGDM's superior performance, consistently outperforming state-of-the-art methods across multiple evaluation metrics. Notably, RGGDM achieves a Peak Signal-to-Noise Ratio (PSNR) of 25.48 dB and an Underwater Image Quality Measure (UIQM) of 4.37 on the UIEB dataset. Furthermore, enhanced images show substantial improvements in downstream tasks such as SIFT feature matching, with an average increase of 132.19\% in matching points. These results underscore RGGDM's potential in advancing underwater visual perception and its broader implications for marine robotics and environmental monitoring applications.