GAN-Driven Inverse Modelling of EM Backscatter Data for 2-D Dielectric Shape Reconstruction

Reconstructing the shape of dielectric scatterers from electromagnetic (EM) backscatter, commonly referred to as EM imaging, presents significant challenges due to the ill-posed and non-linear nature of the problem. Furthermore, the computational complexity of conventional numerical approaches increases substantially with the growing electrical dimensions of the domain of interest. This study proposes a deep learning (DL)-based framework to address these challenges, offering a computationally efficient solution while simultaneously tackling the inherent non-uniqueness of the ill-posed problem. The proposed framework employs an inverse neural network (INN) architecture comprising a dense layer network to map backscatter EM fields into a lower-dimensional latent space constrained to follow a Gaussian distribution. A pre-trained generator reconstructs the scatterer shapes from this latent representation, while a pre-trained forward neural network (FNN) imposes physical constraints to resolve the non-uniqueness issue. Experimental evaluations demonstrate the effectiveness of the trained INN, achieving a mean binary cross-entropy (BCE) loss of 0.04 and a structural similarity index (SSI) of 0.94 on test data. The findings underscore the potential of DL-based approaches to overcome existing limitations in EM imaging and contribute to advancements in computationally efficient scatterer reconstruction techniques.