PKGD: Physics Knowledge-Guided Diffusion Model for Underwater Image Enhancement

Abstract

Underwater images often suffer from color distortion, low contrast, and blurred details due to light absorption and scattering in water, posing significant challenges to underwater vision tasks. Many deep learning-based methods overlook the complex underwater optical imaging process, thereby reducing interpretability and enhancement performance. In addition, light scattering caused by suspended particles can lead to image blurring, making detail enhancement more difficult. To address the above issues, we propose a novel physics knowledge-guided diffusion (PKGD) model that integrates physical priors into the diffusion process to effectively enhance underwater image quality. Specifically, PKGD consists of high-frequency prior generation (HPG) module, underwater imaging constraint (UIC) module, and physical perception diffusion model (PPDM). The HPG module integrates original features with multiscale high-frequency information to generate physical priors such as background light and transmission maps. The UIC module combines these priors with the underwater imaging model to produce a preliminary enhanced image as an auxiliary map. Finally, the PPDM considers both the background light and the auxiliary image to guide diffusion, generating images with clearer textures and higher quality. Experimental results validate the effectiveness of the proposed method. The code is available at https://github.com/Yyh-cc/PKGD.git