Under-sampling high quality parallel single-pixel imaging based on Vision-Transformer

High-quality imaging under low sampling rates is a pivotal challenge for advancing the practical application of single-pixel imaging (SPI). Although significant progress has been made by combining deep learning with SPI, traditional imaging schemes remain limited by low coding resolution, resulting in inefficiency and imaging quality bottlenecks. To address these problems, we propose a parallel single-pixel imaging framework based on array spatial light fields. The framework rearranges the same Hadamard matrix into a 4×4 array light field, which realizes multi-channel compression of physical layer target information and efficient acquisition of aliasing detection signals. Simultaneously, a Vision-Transformer-based reconstruction network is designed to resolve global feature dependencies in aliased signals through a multi-head self-attention mechanism, enabling end-to-end reconstruction of high-resolution images directly from detection values. Experimental results demonstrate that, under an ultra-low sampling rate of 0.31%, the proposed method achieves a 16-fold acceleration in data processing speed compared to conventional SPI. This work overcomes the inherent constraints of coding resolution on imaging efficiency. The end-to-end framework exhibits robust high-quality reconstruction capabilities in complex scenarios, providing an innovative solution for high-speed, large-scale imaging demands in biomedical imaging, industrial non-destructive testing, and related fields.