PolaRestNet: Physics-Guided Computational Reconstruction of Skylight Polarization Patterns for Robust Orientation in Extreme Environments

Skylight polarization patterns provide an important orientation cue for bio-inspired navigation in environments where the global navigation satellite system (GNSS) is unavailable. However, field-of-view (FOV) occlusions and sensing noise frequently induce severe non-linear distortions in the polarization field, compromising the reliability of orientation measurements. To address these challenges, this paper presents PolaRestNet, a physics-guided computational reconstruction framework. At the core of the framework is a polarization-directed attention (PDA) module, which captures global structural features under Rayleigh-scattering geometric priors and improves the recovery of distorted polarization patterns. This design effectively mitigates solar-meridian warping and global pattern asymmetry. To improve model generalization, we construct a Multi-scene Comprehensive Polarization Dataset (MCPD) using a physics-based degradation model for training. Experimental results demonstrate that the proposed method achieves high-fidelity reconstruction of skylight polarization patterns and maintains reliable heading estimation under complex conditions. These results significantly improve both orientation accuracy and system robustness for polarization navigation in challenging environments.