A Physics-Assisted Deep-Learning Scheme Based on Globally Perceptive Modules for Electromagnetic Inverse Scattering Problems

The multiple scattering effect is commonly present in electromagnetic (EM) scattering and can serve as a priori information to guide the solution of inverse scattering problems (ISPs). As depicted by the Green’s function, the multiple scattering effect exhibits both global and local features, i.e., the interaction between subscatterers is in principle global but the coupling strength is usually large in the neighboring region and decays with distance. For solving ISPs, the convolutional layer (CL) has been proven to be highly effective as a fundamental building block in learning-based solvers. Through supervised learning, CL-based networks can predict the refined result from the pre-solution of EM parameters. However, due to the limitation of local perception capability, CL can only model the local feature of the multiple scattering effect. To remedy this, this article proposes a globally perceptive (GP) module that cascades a global self-attention layer (GSAL) and a CL to capture both global interaction and local emphasis. To show the effectiveness of the GP module, a deep-learning scheme with global perception (DLSGP) is constructed using a U-shape structure, where the pre-solution is provided by the backpropagation (BP) or domain current (DC) method. Validation results using cross-set data, noisy data, and measured data demonstrate the high accuracy, strong generalization ability, and robustness of the proposed module. The study on the global perception ability further confirms that the proposed module contains more information on the multiple scattering effect. Since this effect is common in ISPs, this fundamental module can be applied to other learning-based solvers.