Physically Consistent Image Augmentation for Deep Learning in Mueller Matrix Polarimetry.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Mueller matrix polarimetry provides crucial data on light-sample interactions.
• Standard data augmentation techniques often fail to preserve polarization properties.
• This limitation hinders deep learning (DL) model performance in polarimetric imaging.

Purpose of the Study:

• To develop a simulation framework for physically consistent data augmentation of Mueller matrices.
• To validate the physical consistency of the proposed augmentations.
• To demonstrate the benefits of physics-informed augmentation for DL in polarimetric imaging.

Main Methods:

• Introduced a novel simulation framework for applying rotations and flips to Mueller matrices.
• Ensured that transformations maintain the inherent polarization information.
• Validated augmentations against real-world data and applied them to a semantic segmentation task.

Main Results:

• Conventional augmentations were shown to produce falsified results on polarimetric data.
• Physics-based augmentations demonstrated physical consistency with real-world captures.
• Semantic segmentation models using these augmentations showed significant improvements in generalization and performance.

Conclusions:

• Physics-informed data augmentation is essential for robust DL in polarimetric imaging.
• The developed framework enhances dataset diversity and reduces overfitting.
• This approach unlocks the potential of DL for polarimetric datasets, particularly those with limited samples.