Exploring Multiple-Incidence Information in Deep Learning Schemes for Inverse Scattering Problems

Recently, many successes have been witnessed in the field of inverse scattering problems (ISPs) with deep learning schemes (DLSs). However, most of the studies focus on the spatial characteristics of interested targets, and little attentions are paid to the information from multiple incidences that can be treated as temporal information measured in a time series. In this work, a multiple-incidence framework (MIF) is introduced by incorporating the information from multiple incidences into an extra dimension of data space, helping to enrich physical knowledge in DLSs. A 3-D convolutional neural network (CNN) structure is further introduced to explore cross-frame information of the data space, which enables simultaneous utilizations of multiple-incidence information. The proposed MIF can be easily adapted into previous DLSs in the literatures, where two widely used DLSs including back-propagation scheme (BPS) and dominant current scheme (DCS) are used to demonstrate the advantages of the proposed framework. It is shown by extensive numerical and experimental examples that MIF enables consistent improvements for BPS and DCS. It is expected that the proposed MIF will also find its applications in real-time dynamic microwave imaging and multiple-frequency imaging by effectively using physical knowledge from temporal and frequency measurements, respectively.