Deep Learning Approach for Microwave Imaging in Broad Frequency Band Based on Physics-Driven Loss and Deep Convolutional V-Net Structure

Abstract

This article proposes a novel deep learning (DL) approach to realize quantitative real-time microwave imaging (MWI) in the extremely broad frequency band. The proposed DL approach is based on the deep convolutional V-net structure, which employs the residual block and deep convolutional operation to improve its generality and performance. To integrate the physics-based prior to DL model, the inverse-forward closed-loop training framework is introduced to compute the training loss, which comprises two fundamental components: 1) the inverse process for computing data-driven loss, which directly quantifies the dissimilarity between the predictions of our proposed V-net and the actual target contrasts and 2) the forward process for computing physics-driven loss, which evaluates the distinctions between the input EM scattered field and the computed EM scattered field derived from the prediction of V-net. Consequently, the proposed DL method can work with excellent accuracy even for heterogeneous and high-contrast targets, only requiring the single-frequency far-field-measured EM scattered field at the arbitrary frequency in the extremely broad frequency band. Moreover, the proposed DL method can present satisfactory robust on the extremely broad frequency band and provide nearly the same excellent inversion performance on totally different frequencies for one target scatterer. Numerical benchmarks illustrate the feasibility of this proposed DL method.