Unrolled Optimization with Deep Learning-based Priors for Phaseless Inverse Scattering Problems

TL;DR

This paper introduces an end-to-end deep learning framework that unrolls optimization with deep priors to effectively solve highly non-linear and ill-posed phaseless inverse scattering problems under strong scattering conditions, outperforming existing methods.

Contribution

It proposes a novel unrolled optimization network incorporating physics-based models and deep priors, learning all parameters automatically for improved imaging in challenging scenarios.

Findings

Outperforms existing methods significantly.

Provides up to 20 times higher validity range.

Effective in indoor Wi-Fi imaging with strong scattering objects.

Abstract

Inverse scattering problems, such as those in electromagnetic imaging using phaseless data (PD-ISPs), involve imaging objects using phaseless measurements of wave scattering. Such inverse problems can be highly non-linear and ill-posed under extremely strong scattering conditions such as when the objects have very high permittivity or are large in size. In this work, we propose an end-to-end reconstruction framework using unrolled optimization with deep priors to solve PD-ISPs under very strong scattering conditions. We incorporate an approximate linear physics-based model into our optimization framework along with a deep learning-based prior and solve the resulting problem using an iterative algorithm which is unfolded into a deep network. This network not only learns data-driven regularization, but also overcomes the shortcomings of approximate linear models and learns non-linear features. More important, unlike existing PD-ISP methods, the proposed framework learns optimum values of all tunable parameters (including multiple regularization parameters) as a part of the framework. Results from simulations and experiments are shown for the use case of indoor imaging using 2.4 GHz phaseless Wi-Fi measurements, where the objects exhibit extremely strong scattering and low-absorption. Results show that the proposed framework outperforms existing model-driven and data-driven techniques by a significant margin and provides up to 20 times higher validity range.