Dual-constraint reconstruction network with semantic-discriminative consistency for radon single pixel imaging

In recent years, single-pixel imaging (SPI) technology has demonstrated remarkable advantages in terahertz imaging, remote sensing, and hyperspectral imaging due to its unique signal acquisition mechanism. Radon SPI has achieved more efficient and robust classification performance due to direct feature extraction in the radon transform domain. However, at low sampling rates, the performance of radon SPI is severely degraded, leading to a challenging quality enhancement task. In particular, generation-based reconstruction may introduce semantic distortions during the reconstruction process. To address the quality deterioration in radon SPI under low sampling conditions, this paper proposes a deep learning-based reconstruction method. First, a simulated dataset containing typical moving targets (birds) is constructed based on radon projection reconstruction principles. Subsequently, a dual-branch generative adversarial network incorporating a semantic latent vector modulation module is designed. By introducing dual-constraint mechanisms of latent space semantic consistency and image discriminative information preservation, the proposed method effectively suppresses hallucination artifacts while significantly enhancing image detail restoration capabilities. Experimental results demonstrate that the proposed approach can achieve high-quality reconstructed images at typical sampling rates of 2%, 5%, and 25%. The dataset and source codes of the proposed method will be made open-source.