Deep learning-boosted concurrent temporal single-pixel imaging

Abstract

Real-time reconstruction of arbitrary binary masks is achieved via deep learning in temporal single-pixel imaging. To this end, one-time readout temporal ghost imaging (OTR-TGI) generates 560 000 one-dimensional training data which use only 25% of the total that is otherwise needed. A record frame rate of 3 frames per second with high fidelity exceeding that of conventional OTR-TGI is achieved with an advantage of generalization to previously unseen data. Such concurrent retrieval of binary pulse sequences holds promise for metrology, signal as well as information processing.