Computational ghost imaging based on an untrained neural network

• A computational ghost imaging method based on deep learning using an untrained neural network (UNNCGI) is proposed. • Without a large set of labeled data for prior training, the untrained neural network can reconstruct the object image by inputting a set of one-dimensional light intensity. • With the process of UNNCGI, this scheme improves the imaging efficiency and will promote the practical applications of ghost imaging. Ghost imaging based on deep learning (DLGI) usually employs a supervised learning strategy, and needs a large set of labeled data to train a neural network. However, in many practical applications, it is difficult to obtain sufficient numbers of labeled data for training and the training process often takes a long time. Thus, a computational ghost imaging method based on deep learning using an untrained neural network (UNNCGI) is proposed. The input to the network is just a set of one-dimensional light intensity values collected by a single-pixel detector and the neural network can be automatically optimized to generate restored images through the interaction between the network and the process of computational ghost imaging. Both simulation and experiment confirm the feasibility of this untrained network. The reconstructed image of UNNCGI has good quality, even at low sampling ratios, which improves the imaging efficiency and will promote the practical applications of ghost imaging.