Efficient Scattering Center Prediction Method for Targets With Coating Defects Through Deep Learning

Abstract

The dielectric coating defects existing on the surface of low detectable targets have nonnegligible impacts on their electromagnetic (EM) scattering characteristics. Due to the large electrical dimension of the target, traditional EM analysis methods are inefficient and cannot meet the requirement of real-time analysis. To address this issue, this article proposed a U-net-based deep neural network (DNN) to perform efficient scattering center (SC) prediction for targets with coating defects from the input 2-D geometric image of it. Furthermore, facing the difficulty of limited training dataset due to the high cost of full-wave numerical simulations, this article proposed a transfer learning method by pretraining the network on a large dataset obtained by the shooting and bouncing ray (SBR) and then fine-tuning it on the target small dataset obtained by the multilevel fast multipole method (MLFMM), which greatly reduced the training difficulty and improved the accuracy and generalization ability of the model. Numerical results are presented to evaluate the performance of the proposed method. It is proven that the proposed method is promising in providing efficient SC prediction in real-time EM analysis scenarios.