Learned Spinning Mask for High-Speed Single-Pixel Imaging

Single-pixel imaging (SPI) holds significant promise for addressing specialized imaging challenges, particularly in scenarios involving unconventional wavelength ranges and low-light conditions. Recent developments in SPI employing a spinning mask have successfully addressed the limitations of traditional modulators like the Digital Micromirror Device (DMD), particularly concerning refresh ratios and operational spectral bands. However, current spinning mask implementations, relying on random patterns or cyclic Hadamard patterns, struggle to achieve rapid and high-fidelity imaging when operating at low sampling ratios. Here we propose to use deep learning to jointly optimize the encoding and decoding scheme for spinning mask-based SPI. On the encoding side, a cyclic mask, optimized by the convolutional layer, is meticulously crafted to modulate the input object. On the decoding side, the object image is reconstructed from the modulated intensity fluctuations employing a lightweight neural network infused with the image formation model. Our method demonstrates the potential to achieve remarkable imaging results, generating 71×73-pixel images using only a 4% sampling ratio while maintaining a 2.4MHz modulation ratio, yielding image recording speeds surpassing 12KHz. The proposed method dramatically improves the imaging efficiency of SPI, thereby accelerating the practical utilization of SPI in domains such as specialized wavelength imaging and high-speed imaging.