Dual-Module NMM-IEM Machine Learning for Fast Electromagnetic Inversion of Inhomogeneous Scatterers With High Contrasts and Large Electrical Dimensions

Abstract

A dual-module machine learning scheme is proposed to reconstruct inhomogeneous scatterers with high contrasts and large electrical dimensions. The first nonlinear mapping module (NMM) is an extreme learning machine (ELM), which is used to convert the measured scattered fields at the receiver arrays into the preliminary images of the scatterers. The second image-enhancing module (IEM) is a convolutional neural network (CNN), which is used to refine further the images from the NMM to obtain high-accuracy pixel-based model parameter distribution in the inversion domain. Compared with the traditional approximate methods such as backpropagation, the NMM-IEM machine learning can produce the preliminary image with a much higher accuracy but the unknown weight matrices of the ELM are only solved once during training. Hence, the IEM connected to the NMM has a simple architecture and can be trained at a rather low cost. The performance of the proposed dual-module NMM-IEM scheme and the conventional variational Born iterative method is compared in terms of inversion of scatterers with different electrical sizes and contrasts. Meanwhile, the NMM-IEM is also assessed for the inversion of scatterers with high contrasts and large electrical dimensions and experimental data. Finally, the NMM-IEM is compared with the CNNs used in the previous works.