Reconstruction of Two-Dimensional Dielectric Objects with the Conditional Generative Adversarial Network

Abstract

In recent years, deep learning technology is becoming an increasingly important tool to solve inverse scattering problems (ISPs). Due to the existence of multiple scattering, ISPs are highly nonlinear. Traditional methods to solve such problems require a lot of computing costs and are often time-consuming. The traditional Convolution neural network (CNN) is not very accurate. In addition, it can only be used to solve specific targets. In order to improve its accuracy and generalization ability, this paper proposes a deep learning method to improve the solution of ISPs, which combines the conditional general adversarial network (CGAN) with CNN. CGAN is an improvement based on GAN. By adding additional condition information to the generator and discriminator of the original GAN, the unsupervised learning is transformed into supervised learning, and the condition generation model is realized. This paper gives the steps of this method, which divides the model reconstruction into two steps. In the first step, CNN is used to retrieve the initial contrast (dielectric constant) of the scatterer from the measured scattering data to obtain the rough image of the scatterer. In the second step, CGAN is used to enhance the resolution of the rough image. The reconstruction results were better than those without CGAN.