Deep internal learning-based video compressive sensing for the identification of high-frequency structural dynamic characteristics using full-field vision methods

• A self-supervised deep fully-connected network is proposed for video compressive sensing and spatiotemporal super-resolution. • An internal training set generation strategy eliminates external dataset dependence and manual labeling. • The network mitigates motion blur and aliasing, enabling vision-based highfrequency dynamic response recovery. • The proposed method is validated through simulations on a physics-based graphics model and field tests on an actual bridge tower. In computer vision-based structural health monitoring (SHM), spatial and temporal sensor sampling rates are often limited by data storage and transmission requirements, falling below Nyquist rates and restricting high-frequency monitoring. While video compressive sensing partially mitigates these challenges, inadequate temporal modeling during dataset preparation and training hinders the reconstruction of high-frame-rate videos. This paper proposes a deep fully-connected 3D convolutional neural network (DFC-3DNet) for spatiotemporal super-resolution video reconstruction. DFC-3DNet comprises two components: an internal training set generation network and a reconstruction network with one fully-connected and eight 3D convolutional layers. Without requiring external datasets, DFC-3DNet extracts training example pairs directly from the input video, leveraging the strong spatiotemporal recurrence of small video patches for self-supervised training specific to the target structure. Extensive simulations based on a physics-based graphical model (PBGM) validate its superior reconstruction performance, and field applications further demonstrate its potential to improve both the temporal and spatial sampling rates of monitoring videos.