Polarization-guided diffusion prior for eyeglass reflection removal
Eyeglass reflection severely degrades facial feature visibility in video conferencing and facial recognition, where the captured image is a superposition of transmission and reflection layers. Existing polarization-based reflection removal methods depend on large-scale paired polarization datasets, limiting generalization to unseen lighting conditions. This work introduces PDPrior, an untrained polarization-guided diffusion prior that requires no training data and no ground-truth reflection-free images. PDPrior leverages the generative prior of a diffusion model and incorporates polarization information as guidance to control the generation process. The reflection diffusion model uses the degree of linear polarization (DoLP) to preliminarily identify reflection regions and exploits the diffusion prior of progressively darkening facial content, enabling focus on reflective areas. The transmission coefficient computed from DoLP guides transmission image generation via the physical forward model of reflection formation. During each sampling step, reflection and transmission variables are alternately updated through gradient descent based solely on the test sample, conferring adaptability to complex lighting and diverse scenes. Real-world eyeglass reflection images were collected using a division-of-focal-plane polarization camera under various indoor and outdoor lighting environments. Experimental results demonstrate that PDPrior effectively removes eyeglass reflection, producing high-fidelity face reconstructions with no visible artifacts and achieving more robust face image quality assessment scores for recognition performance. The framework generalizes to window photography, showcase displays, and driver monitoring.