Fast Electromagnetic Inversion of Inhomogeneous Scatterers Embedded in Layered Media by Born Approximation and 3-D U-Net

摘要

This letter presents a 3-D electromagnetic inversion method based on the Born approximation (BA) and a convolutional neural network (CNN), the 3-D U-Net. In the training stage, the BA is first used to obtain the preliminary 3-D images of a series of homogeneous scatterers with regular shapes that are further improved by the Monte Carlo method. Then, these images are used to train the 3-D U-Net. In the testing stage, inhomogeneous scatterers with complex shapes are reconstructed by both the trained 3-D U-Net and the traditional iterative method, variational Born iteration method (VBIM). Their performance is evaluated and compared.