Abstract:Polarization image restoration is a critical prerequisite for ensuring the orientation accuracy of polarized light navigation. Most existing studies focus on the influence and restoration of polarization information caused by cloud occlusion, while insufficient attention has been paid to the complete loss of polarization information resulting from opaque hard occlusions such as buildings. Traditional image inpainting methods tend to suffer from edge blurring and texture distortion in hard occlusion scenarios. To address these problems, this paper proposes a sky polarization image inpainting method based on Generative Adversarial Network (GAN). The method adopts a context discriminator that takes both global and local features into account, integrates mean square error (MSE) reconstruction loss, perceptual loss and edge fidelity loss during training, and adopts a step-by-step training strategy to improve the optimization effect of the generator. Experimental results show that under building occlusion conditions, the proposed method can restore polarization images with rich texture details and natural and smooth transitions. The mean heading angle error is reduced from 4.12° to 1.46°, and the directional positioning accuracy is improved by approximately 64.7%. This method can effectively recover the sky polarization features in regions occluded by buildings, providing reliable image preprocessing support for bio-inspired polarized light navigation in complex occlusion environments.