Physics-informed augmentation for deep learning on Mueller matrix polarimetric data

Mueller matrix polarimetry captures rich polarization information but poses challenges for deep learning due to the limited availability of sufficiently large datasets, making effective image data augmentation especially critical. This work introduces a physics-informed data augmentation framework that applies only physically consistent rotations and flips to Mueller matrix images, preserving polarization fidelity. We validate the method through comparisons with real measurements and demonstrate improved model generalization in a semantic segmentation task. This approach enables more robust machine learning on polarimetric datasets, particularly in low-data scenarios, and highlights the importance of embedding physical principles into computational pipelines. The PyTorch implementation is openly available.