Conditional LS-GAN Based Skylight Polarization Image Restoration and Application in Meridian Localization

Abstract

Skylight polarization images (SPIs) contain crucial spatial information that can be used for navigation purposes. Under most circumstances, the quality of the images becomes a major concern, especially when there is blocking between the perception equipment and the sky. This paper introduces a deep learning-based methodology for restoring SPIs and utilizing the restored images for navigation. First, an adversarial training paradigm is adopted to restore the SPIs collected in a severe blocking environment. We utilize a self-developed simulation method that uses solar altitude and azimuth angle to generate ground truth, and no prior knowledge of the masking information for noise is used. Second, we show how to locate the meridian based on the restored SPIs using a residual neural network. In experiments, we demonstrate the superiority of the proposed model in restoring SPIs, and the enhanced meridian localization precision by using the restored SPIs.