Translation-Equivalence-Based Unsupervised Ghost Imaging

Ghost imaging (GI), as a burgeoning imaging technology, has gained a lot of attention in the past few years. To achieve better reconstructed image quality at low sampling rates, deep learning is introduced to GI. While supervised learning is widely used in GI, unsupervised learning also deserves to be studied in depth as it does not need real labeled images to train a network. Following this idea, we proposed a translation equivalence ghost imaging (TEGI) method only based on unlabeled bucket signals. A new loss function is designed to do unsupervised learning. It is demonstrated that our TEGI method can reconstruct target images with better quality than the differential ghost imaging method at low sampling rates.