Enhancing Low-Visibility Images for Operational Environments: Self-Supervised Learning Under Physical Model Guidance

As the demand for underwater operations grows, the need for high-quality underwater imagery in intelligent applications has become increasingly critical. Underwater environments inherently suffer from light absorption and scattering, resulting in images with blue–green hues, blurriness, and low brightness, which pose significant challenges for image enhancement. At the same time, the lack of clear reference images further complicates the enhancement process. To address these issues, we present a self-supervised learning framework guided by an underwater imaging model to enhance image quality. By inverting the underwater imaging model, we decompose the enhancement problem into the acquisition of prior knowledge and parameter estimation. Specifically, we utilize an existing dark channel prior estimation model to estimate the background light prior, while a self-supervised model learns the transmission-related coefficient map, which captures complex, spatially variant degradation. We propose the novel self-supervised framework that, crucially, does not rely on paired clear and degraded images. Instead, it optimizes the model using carefully designed loss functions that leverage intrinsic properties of the degraded image itself. We conducted qualitative and quantitative analyses on public datasets, where our model demonstrated superior performance compared to existing methods. Furthermore, by collecting operational environmental images, we validated the model's effectiveness in operational environments, consistently outperforming other models.