Decoupled sensing matrices with swin transformer for high speed and resolution imaging

High speed and resolution imaging is inherently associated with severe data acquisition burden. To deal with the issue, Spatiotemporal Compressive Imaging (STCI) is a powerful paradigm. However, high compression ratios (CRs) arising from dual-dimensional compression in STCI induce intricate spatiotemporal information coupling, posing formidable challenges for the performance of backend reconstruction algorithms, which limits the development of STCI. To address this limitation, we propose a comprehensive and practical framework for Spatial-Temporal-Decoupled Compressive Imaging (STDCI). Our approach introduces an encoding strategy that decouples a sensing matrix using two masks with distinct spatial scales. Furthermore, we design a reconstruction network, the Spatiotemporal Swin Transformer (STST), based on the Video Swin Transformer (VST) architecture. The proposed method successfully reconstructs high spatiotemporal resolution images from compressive measurements with a total spatiotemporal CR 128:1 (including a temporal CR 8:1 and a spatial CR 16:1). Both numerical and optical experiments demonstrate the superior performance of our integrated encoding-reconstruction framework.