Electromagnetic Inverse Scattering With an Untrained SOM-Net

Abstract

Physics-inspired deep learning (DL) methods become a research hotspot in solving inverse scattering problems (ISPs) due to the advantages of imaging quality and speed. However, the generalization ability is still the main bottleneck of DL-based ISP methods. In this article, an untrained SOM-Net (called uSOM-Net, where SOM stands for the subspace-based optimization), which employs our newly proposed SOM-Net as the backbone neural network, is introduced to solve the ISP without training and testing processes. The uSOM-Net not only achieves high reconstruction quality comparable to existing DL-based methods but also owns similar adaptability as the traditional iterative ones. Specifically, the uSOM-Net parameterizes all physical variables in the Lippmann–Schwinger equation by network weights. To overcome the ill-posedness, the uSOM-Net updates the unknown weights with physical losses defined from both the data and the state equations. Besides, prior information from accessible scatterers can also be easily incorporated into the uSOM-Net by adding extra physical constraints with few prior labeled data. The proposed uSOM-Net is verified with both synthetic and experimental examples to demonstrate its superiority over existing ones. The uSOM-Net intends to bridge the gap between traditional iterative ISP methods and DL ones, which introduces a new way of solving ISPs.