Abstract

In video conferencing and facial recognition applications, eyeglass reflection often obscures critical facial features. Polarization information helps disentangle reflection and transmission components. Existing polarization reflection removal approaches rely on large-scale paired polarization datasets, making them hard to generalize to the scenes under unseen lighting conditions.

To address these challenges, we propose PDPrior, an untrained polarization-guided diffusion prior for eyeglass reflection removal. The PDPrior requires no training data, and no ground-truth reflection-free images. It operates solely on polarization observations collected at test time to achieve artifact-free, high-fidelity reflection removal. The PDPrior leverages the generative prior of a diffusion model and introduces polarization as guidance to control the generation process.

The self-supervised loss based on the forward model is employed to alternately update the reflection and transmission variables, endowing the generative model with physical interpretability. Extensive experimental results validate the effectiveness and robustness of the method on eyeglass reflection data captured in both indoor and outdoor lighting conditions, achieving higher face image quality assessment scores for recognition performance compared to state-of-the-art methods.