Hybrid multi-head physics-informed neural network for depth estimation in terahertz imaging

Abstract

Terahertz (THz) imaging is a topic in the field of optics, that is intensively investigated not least due to its potential for recording three-dimensional (3D) images, useful e.g., for the detection of hidden objects, nondestructive testing, and radar-like imaging in conjunction with automotive systems. Depth information retrieval is a key factor to recover the three-dimensional shape of objects. Impressive results for depth determination in the visible and infrared spectral range have been demonstrated through deep learning (DL). Among them, most DL methods are merely data-driven, lacking relevant physical priors, which thus requires a large amount of experimental data to train the DL models. However, acquiring large training data in the THz domain is challenging due to the time-consuming data acquisition process and environmental and system stability requirements during this lengthy process. To overcome this limitation, this paper incorporates a complete physical model representing the THz image formation process into a DL neural network(NN). Having addressed phase retrieval and image reconstruction of planar objects in an earlier paper, we focus here on the task to retrieve the distance information of objects. A significant goal of our work is to be able to use the DL NNs without pre-training, eliminating the need for tens of thousands of labeled data. Through experimental validation, we demonstrate that by providing diffraction patterns of planar objects, with their upper and lower halves sequentially masked to overcome the trapping of the NN's computational iterations in local minima, the proposed physics-informed NN can automatically reconstruct the depth of the object through interaction between the NN and the physical model. Compared to traditional DL methods and back-propagation methods, our approach not only reduces data dependency and operational costs but also improves imaging speed and stability. The obtained results also represent the initial steps towards achieving fast holographic THz imaging using reference-free beams and low-cost power detection.