A Current-Coding Residual Structure for Solving Electromagnetic Inverse Scattering Problems in Inhomogeneous Background Media

This study proposes a current-coding residual (CCR) structure of deep learning (DL) to solve electromagnetic inverse scattering (EMIS) problems in media with inhomogeneous background structures. The CCR framework combines an encoder-decoder architecture with a residual iterative structure and a current-assisted operator matrix to regulate network parameter learning. By embedding this designed matrix, the CCR imposes constraints on the structural and weight distributions during contrast (permittivity) reconstruction, leveraging the physics-informed interaction on scatterers as prior knowledge, which ensures faster convergence and more stable inversion results. This method focuses on the reconstruction of both the real and imaginary parts of permittivity, making it suitable for lossy scatterers. Special consideration was given to scenarios with different background in the design of numerical experiments, demonstrating the method’s effectiveness in reconstructing high-contrast targets under complex and noisy conditions. Compared to other network-based methods, this approach also offers the advantage of lower computational resource consumption. Finally, its robustness and accuracy were validated using experimental data from the Fresnel Institute.