BDMUIE: Underwater image enhancement based on Bayesian diffusion model

Existing deep learning-based underwater image enhancement (UIE) methods primarily rely on data-driven approaches that map underwater images to their enhanced counterparts. However, these methods often neglect the role of prior information in improving image quality. Such prior information, derived from natural domain images with balanced color distributions and rich structural details, can guide the enhancement process by reducing the domain gap between underwater and natural images. To address this limitation, we introduce BDMUIE, an innovative UIE framework based on Bayesian diffusion models. By incorporating prior knowledge from natural domain images, BDMUIE modifies the reverse diffusion process to effectively reduce the domain gap between underwater and natural images. A novel multi-scale fusion mechanism is introduced, uniquely leveraging encoder features across diffusion branches. This mechanism seamlessly integrates prior information from natural domain images with features from the underwater enhancement branch, resulting in improved detail preservation and color restoration. Additionally, the ECAG block enhances the model's ability to remove noise while supporting the fusion process to retain structural and textural details. Extensive experiments on widely used underwater image datasets, as well as application experiments, demonstrate that BDMUIE achieves superior performance compared to state-of-the-art methods, with significant improvements in both visual quality and quantitative metrics, including PSNR and SSIM.