Multiple Backgrounds in Data-to-Image Deep Learning Schemes for Microwave Imaging

We present a form of dataset augmentation for microwave imaging deep learning schemes that leverages the relationship between the assumed background medium and the physics of the underlying scattering problem. Given a synthetically collected training set we can adopt any of the targets in the set as a background medium by simply adjusting the data accordingly without any additional forward solutions to the scattering problem. Choosing different targets as backgrounds expands the training set size. Results from a trained data-to-image network on such a multi-background augmented dataset demonstrate that we can reproduce the same target relative to many different backgrounds, providing a path towards uncertainty quantification and/or local model explainability.