Fast 3-D Electromagnetic Full-Wave Inversion of Dielectric Anisotropic Objects Based on ResU-Net Enhanced by Variational Born Iterative Method

Abstract

In this article, a novel artificial neural network named residual U-Net (ResU-Net) is proposed to directly reconstruct 3-D dielectric anisotropic objects from scattered electromagnetic field data recorded at the receiver array. ResU-Net has the same framework as that of U-Net but the convolution kernels are replaced with residual kernels. Meanwhile, the squeeze-and-excitation (SE) operation is added to enable information interaction among different channels and further improve prediction accuracy. ResU-Net is trained by thousands of 3-D homogeneous dielectric anisotropic handwritten digits and the corresponding synthesized scattered field data. In the online prediction, ResU-Net can invert multiple anisotropic model parameters of homogeneous 3-D objects instantaneously. For an inhomogeneous object or multiple homogeneous objects, ResU-Net provides good initial profiles which are fed into the following variational Born iterative method (VBIM) full-wave inversion solver. In addition, the VBIM is implemented in a restricted domain instead of the whole 3-D inversion domain to save computational cost. Numerical experiments show that compared with the traditional iterative solver, such as VBIM, the proposed ResU-Net or the hybrid method can not only achieve higher reconstruction accuracy but also accomplish the multiparametric 3-D inversion in a much faster way.