Polarization Navigation Method Based on Deep Learning and Multi-Scale Patch Restoration in Disturbed Environments

To address the issue of increased heading angle calculation errors in atmospheric polarization navigation caused by occlusions (e.g., buildings and vegetation) in disturbed environments, this study proposes a robust framework integrating a DSCA-Unet (DSC-MCA-Unet) deep neural network with multi-scale image reconstruction. It achieves precise segmentation of the occluded region and high-quality restoration of polarization information. Building upon the Unet architecture, we enhance feature extraction through the Dynamic Snake Convolution (DSC) and Multidimensional Collaborative Attention (MCA), achieving a 3.98% improvement in mean Intersection over Union (MIoU) for complex occlusion segmentation tasks. Multiscale recovery strategy, optimized priority calculation and patch matching mechanism are introduced into the traditional patch algorithm to design a multiscale image reconstruction algorithm, which is used to improve the recovery quality of polarized images. The mask from the DSCA-Unet segmentation is combined with the Angle of E vector (AOE) image containing the anomalous pixels for polarization pattern reconstruction, followed by heading angle calculation using restored polarization data. Simulation and measurement experiments demonstrate the proposed framework successfully eliminates the noise caused by occlusions in the polarization information and accurately restores the abnormal regions. The Mean Absolute Error (MAE) of the corrected course angle reaches 0.88°, representing a 79% reduction compared to the direct computation using corrupted polarization data. This study provides theoretical methods and technical support for solving the reliability problem of atmospheric polarization navigation in complex environments.