Reconstruction of partially occluded objects with a physics-driven self-training neural network

TL;DR

This paper introduces a physics-informed neural network approach for reconstructing occluded objects in terahertz holography, leveraging self-training with diffraction data to improve accuracy and noise reduction in 3D imaging.

Contribution

It presents a novel self-training neural network method that incorporates physics principles for improved object reconstruction in THz holography, especially with limited data.

Findings

Effective noise reduction demonstrated

Successful validation with simulated and experimental data

Enhanced 3D imaging capabilities in terahertz systems

Abstract

This study proposes a novel approach utilizing a physics-informed deep learning (DL) algorithm to reconstruct occluded objects in a terahertz (THz) holographic system. Taking the angular spectrum theory as prior knowledge, we generate a dataset consisting of a series of diffraction patterns that contain information about the objects. This dataset, combined with unlabeled data measured from experiments, are used for the self-training of a physics-informed neural network (NN). During the training process, the neural network iteratively predicts the outcomes of the unlabeled data and reincorporates these results back into the training set. This recursive strategy not only reduces noise but also minimizes mutual interference during object reconstruction, demonstrating its effectiveness even in data-scarce situations. The method has been validated with both simulated and experimental data, showcasing its significant potential to advance the field of terahertz three-dimensional (3D) imaging. Additionally, it sets a new benchmark for rapid, reference-free, and cost-effective power detection.