A multi-branch deep convolutional fusion architecture for 3D microwave inverse scattering: stored grain application

Abstract

In this paper, a multi-branch deep convolutional fusion architecture is proposed to solve electromagnetic inverse scattering problems. The conventional methods for solving inverse problems face various challenges, including strong ill-conditioning, expensive computational cost, and unavoidable intrinsic nonlinearity. To overcome these difficulties, we designed a novel multi-branch convolutional neural network (CNN) to reconstruct the 3D images of the moisture distribution in stored grain. Inspired by objective-function techniques for solving the electromagnetic inverse scattering problems, the proposed CNN architecture takes in the scattered-field data and prior information to produce 3D images of the moisture content. With the aim of using inputs of different formats, i.e., a complex-valued vector of scattered-field data and a 3D image of the background moisture distribution as prior information, we propose a multi-branch architecture consisting of decoder-only, and encoder–decoder, convolutional branches. The two branches are later fused to produce the final reconstructed 3D image. To train the CNN, we use the true numerical grain moisture distribution image, which were synthetically generated. The reconstructed moisture distribution images produced by the proposed CNN show that the network is not only able to reconstruct the 3D moisture distribution images directly from measured scattered-field data for high contrast objects-of-interest, but it also achieves a higher imaging quality compared with traditional inversion techniques in microwave imaging. Quantitative evaluations are reported using receiver operating characteristics curves for the hotspot detectability and RMS error. The proposed approach opens a novel path for the deep learning-based real-time quantitative microwave imaging.