Single-Pixel Imaging Based on Deep Learning Enhanced Singular Value Decomposition.

Area of Science:

• Computational Imaging
• Machine Learning Applications
• Optical Sensing

Background:

• Single-pixel imaging (SPI) traditionally relies on predefined patterns like Hadamard sequences.
• Deep learning methods, such as deep convolutional autoencoders, have shown promise in SPI.
• Existing SPI methods face challenges in reconstruction quality, especially at low sampling ratios.

Purpose of the Study:

• To propose and demonstrate a novel single-pixel imaging method.
• To enhance image reconstruction quality and robustness in SPI systems.
• To explore the potential of the method for applications beyond the visible spectrum.

Main Methods:

• Development of a single-pixel imaging method integrating deep learning with singular value decomposition (SVD).
• Theoretical framework and experimental implementation of the deep learning-enhanced SVD (DL-SVD) method.
• Comparison with conventional Hadamard patterns and deep convolutional autoencoder (DCAE) network methods.

Main Results:

• The DL-SVD method achieves superior image reconstruction quality, particularly at low sampling ratios (down to 3.12%).
• The proposed approach requires fewer measurements or shorter acquisition times for comparable image quality.
• Demonstrated enhanced anti-noise performance and improved generalizability to untrained targets.

Conclusions:

• The deep learning-enhanced SVD method offers significant advantages over conventional SPI techniques.
• The method provides robust and high-quality image reconstruction under challenging conditions (low sampling, noise).
• Potential for broad applications in single-pixel imaging, including non-visible spectral ranges.