Compressive high-speed imaging system and reconstruction algorithm

Abstract

High-speed imaging plays a vital role in fields such as biomedical diagnostics, combustion analysis, and astronomical observation, where phenomena evolve on microsecond or even nanosecond scales. Traditional imaging systems, constrained by the Shannon–Nyquist sampling theorem, require high-bandwidth sensors and large data storage, limiting scalability and cost-efficiency. Compressive sensing (CS) offers a paradigm shift by enabling the recovery of high dimensional signals from significantly fewer measurements. Applied to video capture, CS facilitates the design of snapshot imaging systems that reconstruct multiple frames from a single coded measurement. This thesis presents a novel hardware architecture: the Compressive Coded Rotating Mirror (CCRM) system, capable of achieving up to 1.4 million frames per second. The system employs a motor-driven rotating mirror in combination with static binary masks to spatially encode successive frames across different pixel columns, enabling high frame counts without relying on expensive digital mirror devices (DMD). To complement the hardware, we develop a reconstruction framework that integrates classical optimization (e.g., Generalized Alternating Projection and ADMM) with domain specific priors. A key innovation is the foreground–background decomposition strategy, which enhances reconstruction by leveraging temporal redundancy. The use of Total Variation regularization and plug-and-play denoisers further improves robustness across various dynamic scenes. Extensive experiments, including synthetic simulations and real-world capture of fast phenomena, validate the system’s effectiveness. The proposed CCRM system offers a scalable and cost-efficient solution for ultra-high-speed video acquisition.