URNet: High-quality single-pixel imaging with untrained reconstruction network

•A single-pixel imaging (SPI) method using an untrained reconstruction network (URNet) is proposed.•Only a single 1D data is needed to feed the URNet, and the method can automatically be optimized and eventually retrieve the 2D image without training tens of thousands of labeled data.•This method can solve the limitation of generalization ability and interpretability in the supervised strategy-based deep learning SPI method.•Reconstructed results show that the proposed method outperforms other widespread methods in terms of visual quality and noise immunity especially in the case of very low sampling rate. High quality image reconstruction method is an important guarantee for the practical application of single-pixel imaging (SPI). The supervised strategy-based deep learning SPI method requires manual labeling of thousands of training sets to optimize the network model, which needs to take several days or even months to label such data. In addition, generalization ability and interpretability limit the application of the supervised strategy-based deep learning SPI method. According to this, a SPI method using an untrained reconstruction network (URNet) is proposed. In this scheme, only a single 1D data collected by the photodiode is needed to feed the URNet, and the network can automatically be optimized and eventually retrieve the 2D image without training tens of thousands of labeled data. Reasonable reconstructions indicate that the proposed method outperforms other widespread reconstruction methods in terms of visual quality and noise immunity especially in the case of very low sampling rate, which can further expand the practical application of SPI.