Self-supervised high-pixel-resolution reconstruction for single-pixel imaging via measurement decoupling

Single-pixel imaging (SPI) is constrained by the optical resolving power of the imaging system and the limitations of its sampling mechanism, making high-pixel-resolution (HPR) imaging difficult to achieve and leading to degradation in structural details in reconstructed images. In this work, we propose a self-supervised SPI reconstruction method. The method can directly generate a target HPR image using a neural network, and then downsample it to the measurement-pixel-resolution (MPR) grid for optimization. By imposing consistency constraints on the network in the measurement domain via the SPI physical model, it enables convergent reconstruction of HPR images without external supervision. Based on this framework, the dependency between physical MPR grid and the target reconstruction space is decoupled, allowing the algorithm to recover HPR images from low-resolution MPR data and alleviating the information representation limitation caused by the constrained MPR dimension. Simulation and experimental results demonstrate that the proposed method achieves stable reconstruction performance, preserves fine structural details effectively, and outperforms several existing methods in both visual quality and quantitative metrics, providing an effective imaging solution for applications such as biomedical imaging.