A Physics-Induced Deep Learning Scheme for Electromagnetic Inverse Scattering
Zeyang Wu, Yuexing Peng, Peng Wang, Wenbo Wang, Wei Xiang
Abstract
This article studies the full-wave electromagnetic inverse scattering (EMIS) problem that aims to retrieve the permittivities of dielectric scatterers based on the knowledge of measured scattered fields. A physics-induced deep learning (DL) scheme termed contrastive learning-based subspace optimization and semantic segmentation-assisted reconstruction (CLSO-SSR) is proposed to solve this problem. The proposed CLSO-SSR model is novel in aspects: 1) a contrastive learning network to provide an initial estimate of the contrast source so that both the probability of the local optima and the time consumption of the inversion process are greatly reduced and 2) a semantic segmentation network with an attention mechanism to mitigate the nonlinearity caused by the multiple scattering effects. Extensive numerical experiments are carried out to show that the proposed CLSO-SSR scheme can significantly improve upon the permittivity reconstruction performance as opposed to existing alternatives.