Physics-informed Unsupervised Deep Learning Framework for Solving Full-Wave Inverse Scattering Problems

Abstract

A classical inverse scattering problem (ISP) aims at retrieving images of permittivity distributions for dielectric scatterers. The strong nonlinearity and ill-posedness of ISPs making canonical inverse methods difficult to produce high quality images for intricate and strong scatterers. Nowadays, many deep learning frameworks for ISPs have sprung up and are proved to possess high-imaging quality. However, they all need paired training samples consisting of massive measured scattered fields and corresponding target images to learn the mapping relationship for inverse imaging. The collection of scattered fields in training dataset mainly depends on full wave simulation such as the method of moments (MoM), which takes a lot of time and computing resource, and even becomes unpractical when the imaging scale is large. Here, we propose a physics-informed unsupervised deep learning framework, termed CSI-GAN, for ISPs. This framework imbeds a whole contrast source inversion (CSI) procedure and an unsupervised generative adversarial network (GAN) where the CSI provides physical constraint to GAN, and GAN adds topological and semantic features to CSI, working together realizing high quality inverse imaging for strong scatterers. The effectiveness of CSI-GAN is verified by simulation tests. To our best knowledge, the proposed CSI-GAN is the first unsupervised deep learning framework for inverse scattering imaging. Compared with previous deep learning frameworks, measured scattered fields are no longer needed in our training dataset, enormously reducing the cost and limitation for deep-learning-based ISPs.