Deep Learning Assisted Distorted Born Iterative Method for Solving Electromagnetic Inverse Scattering Problems

Abstract

This paper presents the deep learning-assisted Distorted Born Iterative Method (DBIM) for the permittivity reconstruction of dielectric objects. The inefficiency of DBIM to reconstruct strong scatterers can be overcome if it is supported by Convolutional Neural Network (CNN). A novel approach, cascaded CNN is used to obtain a fine-resolution estimate of the permittivity distribution. The CNN is trained using images consisting of MNIST digits, letters, and circular objects. The proposed model is tested on synthetic data with different signal-to-noise ratios (SNRs) and various contrast profiles. Thereafter, it is verified through experimental data provided by the Institute of Fresnel, France. Reconstruction results show that the proposed inversion method outperforms the conventional DBIM method in terms of accuracy and convergence rate.