Learning-Based Quantitative Microwave Imaging with a Hybrid Input Scheme

In this letter, a learning-based inversion method with a hybrid input scheme is proposed to solve quantitative microwave imaging (MI) problems. The high-resolution dielectric targets are reconstructed by convolutional neural network (CNN) with a hybrid input scheme. The qualitative direct sampling method (DSM) is utilized to provide the spatial information, while the quantitative back propagation (BP) is employed to get the preliminary constitutive parameters of the unknown target. The hybrid input scheme, defined as an inner product of DSM and BP results (shorten as a DSM-BP scheme), significantly improves the reconstruction quality of U-net CNN compared to the BP-only input scheme without additional computational burden. The accuracy and stability of the proposed inversion are verified by both synthetic and experimental data.