Toward Constrained Generative Adversarial Network Architecture for Precision Microwave Imaging

Abstract With the recent advancements in artificial intelligence, there has been a continuous pursuit of efficient solutions for EM imaging problems. Herein, a deep learning approach was presented using generative adversarial network to reconstruct two-dimensional dielectric objects from multi-frequency scattered field amplitudes. An adversarial autoencoder was trained to generate scatterer geometries from a Gaussian-constrained latent space, while a cohesive inverse neural network (INN) integrated dense layers, the trained generator, and a forward neural network. The images reconstructed at the INN output were validated through the forward model, addressing non-uniqueness. The INN achieved a mean binary cross-entropy loss of 0.04 and a structural similarity index of 0.94. The study not only demonstrates a significant reduction in computational load, but also marks an improved accuracy over traditional objective-function based methods. It contributes both to the fields of machine learning and EM imaging by offering a real-time quantitative imaging approach. The obtained results would open up new avenues for radio-frequency inverse imaging.