Learning-based high-speed single-pixel imaging using a cyclic random mask

Single-pixel imaging (SPI) is an advanced computational imaging technique that employs a simple bucket detector to capture object images without raster scanning. This method offers advantages such as low cost, high sensitivity, and suitability for imaging in low-light environments and specialized wavebands. However, SPI inherently suffers from a limitation in imaging speed due to the need to acquire intensity fluctuation signals under a large number of spatially modulated patterns. Here, we tackle this challenge by developing a high-speed optical modulation system and an advanced reconstruction algorithm, which together enhance the refresh rate of the optical modulation process while reducing the required sampling ratios, thereby enabling high-speed SPI. Specifically, on the hardware side, we implement a spinning disk modulation scheme with cyclic random patterns coded onto the disk, achieving a modulation refresh rate of 1 MHz. On the algorithmic side, we propose a physics-enhanced deep learning framework combined with a lightweight neural network, LiteUNet, which reduces the required sampling rate to 10%. By combining these innovations, we experimentally demonstrate high-speed SPI at 1926 fps with a spatial resolution of 71 × 73 pixels. This work offers an effective solution to address the imaging speed bottleneck in SPI, paving the way for its practical applications in fields such as microscopy and remote sensing.