Part-based image-loop network for single-pixel imaging

In this study, we proposed a self-supervised image-loop neural network (ILNet) with a part-based model for single-pixel imaging (SPI). ILNet employs a part-based model that divides image features into different parts to facilitate finer-grained learning, resulting in improved image details when reconstructing a randomly input 2D signal into a 2D object image. Then, the 2D image generated by ILNet can serve as input for the subsequent iteration to continuous incorporation of prior information to ensure high-quality imaging at low sampling rates. 1D signals collected by the single-pixel detector are used as labels for adaptively optimizing and reconstructing the image. Our results show that the ILNet can reconstruct high-quality images with lower sample rates in un-known free-space and underwater experiments, making it a general framework for incorporating physical models into neural networks and expanding the practical application of SPI.