Differentiable Physical Optics Method for 3-D Reconstruction of Dielectric Objects

Lizhen Yang,Hai Lin
DOI: https://doi.org/10.1109/ap-s/inc-usnc-ursi52054.2024.10685886
2024-01-01
Abstract:This paper presents a novel machine learning framework for 3-D reconstruction of dielectric targets from scattered electromagnetic (EM) field data. Our framework comprises two pivotal components: the shape representation part, which utilizes neural implicit surfaces as the continuous geometric representation of the reconstructed model, and the EM parameters storage part, which employs a fully-connected neural network to store the EM properties of the reconstructed model. Our algorithm takes the incident wave configuration as input and outputs the corresponding scattered field computed by the differentiable physical optics method (DPO). Subsequently, we optimize the geometric shape and the EM parameters by minimizing the residual between the output and the measured scattered field. Our numerical experiments demonstrate promising results in 3- $\mathrm{D}$ reconstruction tasks.
What problem does this paper attempt to address?