Self-supervised underwater image enhancement method using dual attenuation coefficient physical modeling and multi-dimensional feature fusion

Underwater imaging is often degraded by scattering and uneven illumination, which can lead to color distortion, reduced contrast, and loss of fine details. Many existing underwater image enhancement methods employ a single attenuation coefficient to model both direct transmission and backscattering. However, this simplified assumption may not fully capture the different physical attenuation behaviors involved in underwater imaging, which can limit performance under varying water conditions and lighting environments. In this work, we propose an underwater image enhancement framework that combines dual attenuation coefficient physical modeling with prior-guided self-supervised learning. The proposed approach models the attenuation processes of direct light and backscattering separately, aiming to better reflect their distinct physical characteristics. In addition, a multi-dimensional feature fusion strategy is introduced for background light estimation, together with a homology consistency constraint and a region-level perceptual fusion mechanism to improve structural consistency and detail representation. The method is trained in a self-supervised manner without requiring large-scale paired datasets. Experimental results on six benchmark datasets indicate that the proposed approach achieves competitive performance compared with 16 recent underwater image enhancement methods, showing improvements in color restoration, contrast enhancement, and detail preservation. The code is publicly available at: https://github.com/LesterQF/SDFM .