Deep learning-enhanced qualitative microwave Imaging. Rationale and Initial assessment

In this paper, an innovative approach to microwave imaging that combines qualitative imaging and deep learning is presented. The goal is to set a framework for a reliable and user-independent retrieval of the shapes of unknown targets. To this end, the proposed approach exploits an inversion technique known as orthogonality sampling method, which is capable of providing a qualitative estimation of the shape of targets in realtime. The output of the qualitive inversion is processed by a deep learning fully convolutional network called U-Net. U-Net automatically generates binary masks depicting the geometrical properties of the targets, i.e., separates the scattering objects (foreground) from the background. A quantitative assessment of the performance of the processing framework is provided with simulated data to demonstrate the capabilities of the proposed approach.