Enhanced single-pixel imaging base on posterior mean prediction using a score-based diffusion model

In Single-pixel imaging (SPI), reconstructing high-quality images at low measurement rates remains a significant research challenge. However, existing methods face issues such as insufficient image reconstruction quality. This paper proposes a novel reconstruction approach using a score-based diffusion model to address this challenge. First, to effectively apply the score-based diffusion model to SPI, we introduce a measurement likelihood representation grounded in the imaging system. This likelihood term, which incorporates measurement information, serves as a conditional constraint in the reverse diffusion process for image reconstruction. Additionally, we propose a posterior mean estimation network to more accurately predict the posterior mean of the target image, thereby improving imaging quality of SPI. Experimental results demonstrate that our proposed method outperforms the state-of-the-art SPI techniques across various measurement rates. Furthermore, ablation studies and noise robustness tests further validate the effectiveness and practical applicability of the proposed approach.