Physics-Informed Deep Neural Network for Polarimetric Descattering Imaging in Dynamic Cement Dust Environments

Polarimetric descattering imaging has attracted growing interest due to its fundamental physical significance and potential applications. While deep learning has accelerated its development through powerful feature extraction and inference capabilities, existing methods still face limitations in practical scenarios, particularly under dynamic non-uniform scattering conditions such as cement dust environments. To address this, we propose a deep neural network based on the Mueller matrix model that effectively integrates polarization evolution information with deep learning. Specifically, local concentrations of the scattering medium in non-uniform cement dust are characterized by the evolution of the degree of linear polarization (DoLP), which is converted into pixel-wise weight biases to generate customized Mueller matrices adaptable to varying concentrations. The network predicts a pixel-wise dust concentration map and applies the corresponding concentration-specific Mueller matrix to each pixel for polarization-aware dehazing, ensuring physical consistency with Mueller matrix calculus throughout inference. This framework is further enhanced by a physics-constrained optimization loss and multi-scale feature fusion. Experimental results demonstrate the method’s effectiveness and superiority in diverse dynamic non-uniform cement dust environments.