A Lightweight Deep Learning Model for Full-Wave Nonlinear Inverse Scattering Problems

Nowadays, deep learning schemes (DLSs) have gradually become one of the most important tools for solving inverse scattering problems (ISPs). Among DLSs, the dominant current scheme (DCS), which extracts physical features from the dominant components of the induced currents, has shown its successes by simplifying the learning process in solving ISPs. It has shown excellent performance in terms of efficiency and accuracy, but the increasing number of channels in DCS often requires higher computational costs and memory usage. In this paper, a lightweight deep learning model for DCS is proposed to reduce the burden of memories in the training and testing processes of network structure. And extensive tests of the model are conducted, where comparisons with results from the U-Net structure are provided. The comparison results validate its potential application in utilizing DCS under limited resource conditions.