Deep learning based video snapshot compressive imaging of Leidenfrost droplets

Conventional high-speed imaging is constrained by bandwidth, storage, and cost, limiting both temporal resolution and recording duration. We present a Coded Aperture Compressive Temporal Imaging (CACTI) framework that combines optical temporal encoding with deep-learning-based reconstruction to enable low-cost, high-speed video acquisition under strong compression. The system optically encodes twenty sequential frames into a single measurement and reconstructs videos at an equivalent rate of 1200 fps, achieving substantial data reduction while maintaining high fidelity. To validate its performance, the framework was applied to visualize droplet dynamics on heated surfaces, a scenario demanding sub-millisecond temporal precision and fine spatial resolution. Under these conditions, the system clearly captured the complete bouncing and deformation process associated with the Leidenfrost effect. Comparative experiments under identical conditions show that the reconstructed videos exhibit image clarity and temporal continuity comparable to a commercial high-speed camera, while requiring only about one-twentieth of the data throughput per acquisition. These results demonstrate that the CACTI framework provides a scalable, compact, and high-throughput platform for studying rapid thermofluidic phenomena, highlighting the potential of computational imaging to replace costly high-speed hardware for real-time observation of transient physical processes.