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

Area of Science:

• Computational Imaging
• Optics
• Machine Learning

Background:

• Single-pixel imaging (SPI) offers cost-effective, sensitive imaging, ideal for low-light and specialized wavebands.
• A key limitation of SPI is its slow imaging speed due to extensive sampling requirements.
• Overcoming this speed bottleneck is crucial for broader SPI adoption.

Purpose of the Study:

• To develop a high-speed single-pixel imaging system.
• To enhance the refresh rate of optical modulation and reduce sampling ratios.
• To enable practical, high-frame-rate SPI for advanced applications.

Main Methods:

• Implemented a spinning disk modulation system with cyclic random patterns achieving a 1 MHz refresh rate.
• Developed a physics-enhanced deep learning framework utilizing a lightweight neural network (LiteUNet).
• Reduced the required sampling rate to 10% through algorithmic optimization.

Main Results:

• Demonstrated high-speed SPI experimentally at 1926 frames per second (fps).
• Achieved a spatial resolution of 71 × 73 pixels.
• Successfully addressed the inherent speed limitations of SPI.

Conclusions:

• The developed system effectively overcomes the imaging speed bottleneck in SPI.
• This advancement paves the way for practical SPI applications in microscopy and remote sensing.
• Combines hardware innovation with advanced AI for high-performance imaging.