Self-supervised learning for single-pixel imaging via dual-domain constraints

Deep-learning-augmented single-pixel imaging (SPI) provides an efficient solution for target compressive sensing. However, the conventional supervised strategy suffers from laborious training and poor generalization. In this Letter, we report a self-supervised learning method for SPI reconstruction. It introduces dual-domain constraints to integrate the SPI physics model into a neural network. Specifically, in addition to the traditional measurement constraint, an extra transformation constraint is employed to ensure target plane consistency. The transformation constraint uses the invariance of reversible transformation to impose an implicit prior, which avoids the non-uniqueness of measurement constraint. A series of experiments validate that the reported technique realizes self-supervised reconstruction in various complex scenes without any paired data, ground truth, or pre-trained prior. It can tackle the underdetermined degradation and noise, with ∼3.7-dB improvement on the PSNR index compared with the existing method.