Improving the use of Multiple Backgrounds for Deep Learning-Enabled Microwave Imaging

We discuss two possible improvements to a recently presented method of a multiple background approach to uncertainty quantification for machine learning enabled microwave imaging. A microwave imaging training set is partitioned into a set of training targets and a set of artificial backgrounds that are used to create an augmented training set. A single test target is evaluated against each of the artificial backgrounds creating a set of test examples which can be used to create a mean prediction as well as an uncertainty map (standard deviation of the set of predictions). Herein we discuss a limitation of the method as originally presented and two possible means of mitigation. Results show that by leveraging the knowledge of the artificial backgrounds, we can improve the quality both of the reconstructed target, and its corresponding uncertainty map.