Two Steps Electromagnetic Quantitative Inversion Imaging Based on Convolutional Neural Network

This study focuses on tackling the complex issue of 2-D full-wave inverse scattering problem (ISP), which aims to deduce the dielectric properties of a scatterer using collected scattering data. Traditional techniques for addressing inverse problems, such as subspace optimization, contrast source inversion, and the Born iterative method, often encounter challenges like severe ill-conditioning, substantial computational requirements, inherent nonlinearity, and a propensity for converging to incorrect local minima, especially in cases involving materials with high dielectric constants. To navigate these hurdles, we introduce an innovative two-phase method. Initially, a backpropagation scheme is adapted to derive a preliminary approximation of the scatterer's dielectric constant from the scattering data obtained. Subsequently, this preliminary approximation is enhanced through a U-Net model, aiming at improving the quality of the image reconstruction. The effectiveness and precision of our proposed solution have been confirmed through experimental results, suggesting its potential for facilitating real-time quantitative imaging.