Physics-guided multi-scale feature extraction and dual-path refinement network for underwater polarization image descattering

Images captured in turbid underwater environments often suffer from severe backscattering, low contrast, and blurred details, which degrade image quality and affect subsequent visual tasks. To address these issues, we propose MFDP-Net, a physics-guided network for underwater polarization image descattering. Given three polarized images captured at different analyzer angles, the network first employs a multi-scale contextual polarization feature extraction module to capture local details and multi-scale contextual information from different polarized inputs. The extracted features are then fed into a dual-path feature refinement module for further refinement, where two parallel feature paths are used to enhance complementary feature responses and improve the representation of degraded underwater images under complex turbidity conditions. Based on the refined features, the network estimates polarization-related intermediate variables and uses them in the polarization-guided underwater image descattering model to recover a clear image. In addition, a Stokes consistency loss is introduced to constrain the Stokes components of the predicted polarized images to be consistent with those of the input images, encouraging polarization-consistent restoration. Experiments on the self-collected underwater polarization dataset show that MFDP-Net outperforms representative methods in restoration quality and detail preservation.