Advancements in Terahertz Imaging Through Physics-Driven Neural Networks

Recent advancements in terahertz (THz) imaging technology have been made with the introduction of a physics-informed deep learning algorithm aimed at reconstructing partially occluded objects. This approach leverages the angular spectrum theory to generate a dataset of diffraction patterns that encapsulate information about the objects in question. The researchers, Mingjun Xiang, Kai Zhou, Hui Yuan, and Hartmut G. Roskos, utilized both simulated and experimental data to train a self-training neural network (NN). This NN iteratively predicts outcomes from unlabeled data, reintegrating these results into the training set, which enhances the model's ability to reduce noise and minimize interference during object reconstruction.

The method demonstrates significant potential for advancing three-dimensional (3D) imaging in the THz range, particularly in scenarios where data is scarce. The findings indicate that this technique not only improves the accuracy of object reconstruction but also sets a new benchmark for rapid and cost-effective power detection without the need for extensive reference data. The implications of this research could extend to various fields, including materials science and biomedical imaging, where accurate imaging of obscured objects is crucial.

The full paper, titled "Reconstruction of partially occluded objects with a physics-driven self-training neural network," can be accessed on arXiv at arXiv:2408.13066.