A Contrast-Source Inversion-Assisted Attention-Unet for Microwave Imaging

This paper introduces a physics-based intermediate estimate as input to the Attention-Unet (ATTN-Unet) architecture for solving the electromagnetic inverse scattering problem in microwave imaging. This input is calculated from the tenth iteration results of the conventional contrast source inversion (CSI) algorithm, which is referred to as ITER10. This input incorporates more physical domain knowledge than the widely used backpropagation (BP) estimate, which corresponds to the zeroth iteration of the CSI. While the non-iterative BP estimate is popular due to its simplicity, this work demonstrates that using ITER10 enhances reconstruction accuracy without significantly increasing computational cost. For comparison and the validation of choosing ITER10, the performance of the ATTN-Unet is evaluated using estimates from other intermediate CSI iterations, namely ITER5, 20, 30, and 40. Further, the network outputs are the reconstructed relative complex permittivity values (real and imaginary) of an imaged object. The networks are tested using synthetic and experimental datasets. The results show that the ITER10-ATTN-Unet significantly enhances reconstruction accuracy, outperforming both the BP-ATTN-Unet and the conventional CSI method. Furthermore, the results demonstrate that the ITER10-ATTN-Unet achieves a better balance between accuracy and computational cost compared to the other ITER-based models. These findings highlight the effectiveness of ITER10 as a strong alternative to BP in improving neural network reconstructions in microwave imaging.