Hybrid Multi-Head Physics-informed Neural Network for Depth Estimation in Terahertz Imaging

TL;DR

This paper introduces a hybrid physics-informed neural network for depth estimation in terahertz imaging, effectively combining physical models with deep learning to reduce data requirements and improve 3D object reconstruction.

Contribution

It proposes a novel physics-informed neural network that integrates physical models into deep learning for THz depth estimation without pre-training or large datasets.

Findings

Successfully reconstructs object depth using diffraction patterns.

Eliminates the need for extensive labeled training data.

Demonstrates potential for fast, reference-free THz holographic imaging.

Abstract

Terahertz (THz) imaging is one of the hotspots in the field of optics, where the depth information retrieval is a key factor to restore the three-dimensional appearance of objects. Impressive results for depth extraction in visible and infrared wave range have been demonstrated through deep learning (DL). Among them, most DL methods are merely data-driven, lacking relevant physical priors, which thus request for a large amount of experimental data to train the DL models.However, large training data acquirement in the THz domain is challenging due to the requirements of environmental and system stability, as well as the time-consuming data acquisition process. To overcome this limitation, this paper incorporates a complete physical model representing the THz image formation process into traditional DL networks to retrieve the depth information of objects. The most significant advantage is the ability to use it without pre-training, thereby eliminating the need for tens of thousands of labeled data. Through experiments validation, we demonstrate that by providing diffraction patterns of planar objects with their upper and lower halves individually masked, the proposed physics-informed neural network (NN) can automatically optimize and, ultimately, reconstruct the depth of the object through interaction between the NN and a physical model. The obtained results represent the initial steps towards achieving fast holographic THz imaging using reference-free beams and low-cost power detection.