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

We investigate artificial-intelligence-supported in-line holographic imaging with coherent terahertz (THz) radiation. The goal is to reconstruct three-dimensional (3D) scenes from images obtained with detectors that recorded only the power of the radiation and not the phase. This study proposes a novel approach utilizing a physics-informed deep learning (DL) algorithm to reconstruct objects which partially obscure each other. Taking the angular spectrum theory as prior knowledge, we generate a synthetic dataset consisting of a series of diffraction patterns that contain information about the type of objects to be imaged. This dataset, combined with unlabeled data obtained by experiments, are used for the self-training of a physics-informed neural network (NN). During the training process, the NN iteratively predicts images of the objects from the unlabeled dataset and reincorporates these results back into the training set. This recursive strategy includes experimentally recorded images from the studied object class in the NN training, for which the ground truth is unknown. Furthermore, the approach 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 3D THz imaging.