A transformer-embedded learning method for solving full-wave inverse scattering problems

Electromagnetic inverse scattering problems (ISPs) have been constrained by challenges such as nonlinearity and ill-posedness. This paper introduces a learning-based electromagnetic inversion approach without iterative computation dealing with these challenges. A residual block-constructed U-shape network called TransISP is developed to address the dense pixel-level microwave imaging task. Transformer blocks are embedded as the bottleneck of the TransISP to extract the latent features of the high-contrast target. Numerical experimental results demonstrate that TransISP can effectively learn the knowledge mapping the radiative part to the total contrast and outperforms the method without transformers in the quality of inversions.