Non-pre-trained complex-valued neural network-enhanced Fourier single-pixel imaging

Area of Science:

• Computational Imaging
• Optical Physics
• Signal Processing

Background:

• Single-pixel imaging (SPI) offers robust anti-interference in low-light conditions.
• Current SPI methods face limitations in image quality and reconstruction efficiency at low sampling rates.

Purpose of the Study:

• To propose a non-pre-trained complex-valued network-based compressive Fourier SPI technique.
• To enhance reconstructed image quality and reconstruction efficiency under low sampling conditions.

Main Methods:

• Utilized a physical model as prior information for network-driven optimization.
• Developed a complex-valued neural network that does not require pre-training on datasets.
• Employed compressive Fourier single-pixel imaging.

Main Results:

• Achieved a peak signal-to-noise ratio of 23.68 dB at a sampling rate of 0.05 for 64x64 pixel images.
• Demonstrated high-performance image reconstruction at significantly reduced sampling rates.
• Validated the technique through both simulation analysis and experimental results.

Conclusions:

• The proposed complex domain network-based compressive SPI imaging scheme overcomes current limitations.
• This technique paves the way for practical applications of SPI, especially under low sampling conditions.
• Offers a high-performance, efficient solution for computational imaging challenges.