Spatiotemporal learning parallel single-pixel imaging for dual-layer 3D reconstruction through scattering obstruction

Three-dimensional imaging through scattering obstructions is challenged by multi-path signal mixing. Although parallel single-pixel imaging (PSI) demonstrates strong capability in decoupling complex illumination, the large number of required basis patterns becomes a bottleneck for imaging efficiency. In this work, we propose a spatiotemporal learning parallel single-pixel imaging method (STL-PSI) that achieves high-efficiency dual-layer reconstruction through scattering obstructions. The proposed STL-PSI redefines the data acquisition paradigm by designing a two-stage learning framework to achieve single-frame basis spectral coefficient recovery, which replaces the conventional multi-step phase-shifting mechanism. Furthermore, by leveraging the joint constraints of temporal difference characteristics and the spatial phase retrieval model, we establish a physics- and data-driven pipeline to improve the accuracy of spectral coefficient prediction. Experiments validated that STL-PSI reduces the required basis patterns from over 130 to as few as 21, which corresponds to a reduction of approximately 84%, while simultaneously achieving comparable measurement accuracy.