Attention-Unet for Electromagnetic Inverse Scattering Problems in Microwave Imaging

Abstract

Deep convolutional neural networks (CNNs) are investigated to solve inverse scattering problems for microwave imaging (MWI). The conventional approaches for solving inverse problems encounter challenges such as noisy data and high computational costs. Thus, various deep-learning techniques have been proposed recently to tackle these issues. In this article, the attention-Unet (ATTN-Unet) architecture with attention gates (AGs) is implemented for MWI applications. Further, it is compared against the performance of other CNN-based architectures with similar configurations, namely, DCEDnet, Unet, and Unet-Lite. In addition, the Unet-Lite is implemented with AGs, mainly to evaluate the consistency of performance improvement due to AGs. All the networks have been implemented and tested with complex—real and imaginary—inputs and outputs. The inputs are the backpropagation (BP) of the measured scattered fields onto the imaging domain. The outputs are the reconstructed real and imaginary relative complex permittivity values of an object-of-interest (OI). The results from different networks are compared against each other and against the conventional contrast source inversion (CSI) algorithm. The proposed ATTN-Unet is then tested with experimental data from the University of Manitoba (UM) repository. The results show that the implemented deep-learning method produces image reconstructions of better quality with much lesser computational time.