A Deep Learning Architecture for Augmented Shape Reconstruction via Microwave Imaging

In this paper, an innovative microwave imaging approach that combines deep learning techniques and qualitative inversion methods is presented. In particular, the proposed approach is meant for imaging piece-wise homogeneous targets and aims at providing an augmented morphological reconstruction, which not only retrieves the shape of the targets, but also the spatial variations of the permittivity values. Such an information is not displayed by qualitative inversion methods; however it is efficiently encoded in the gradient of the unknown contrast. In particular in this paper, a physics-assisted deep learning technique, where domain knowledge is given in the inputs of a U-Net architecture, is developed. The domain knowledge is provided by the qualitative image of the unknown targets obtained using the orthogonality sampling method, thus allowing the architecture to provide, once trained, a fully automated and real-time prediction. An initial assessment for the approach with synthetic data is provided.