Amplitude/Phase Retrieval for Terahertz Holography With Supervised and Unsupervised Physics-Informed Deep Learning

Most neural networks proposed for computational imaging (CI) in the terahertz (THz) bands require a large amount of experimental data to optimize their weights and biases. However, obtaining a sufficient number of ground-truth images for training is challenging in the THz domain due to the requirements of environmental and system stability, as well as the lengthy data acquisition process. To overcome this limitation, this article proposes novel supervised and unsupervised physics-informed deep learning (DL) methods for amplitude and phase recovery by incorporating angular spectrum diffraction theory as prior knowledge. First, we demonstrate that our unsupervised dual network can predict both amplitude and phase simultaneously, overcoming the limitations of previous studies that could only predict phase objects. This is demonstrated using synthetic 2-D image data as well as measured diffraction images. The advantage of unsupervised DL is its ability to be used directly without labeling by human experts. In addition, we address supervised DL, which is a concept of general applicability. We introduce training with a database set of 2-D images taken in the visible spectra range and numerically modified by us to emulate THz images. This approach allows us to avoid the prohibitively time-consuming collection of a large number of THz-frequency images. Furthermore, we employ a combination method that enhances the sharpness of image edges, improves contrast, and effectively aligns the approach with the ground truth. The results obtained using both approaches represent the initial steps toward fast holographic THz imaging with reference-beam-free, low-cost power detection.