Comprehensive compensation of real-world degradations for robust single-pixel imaging

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Single-pixel imaging (SPI) struggles with image quality due to complex real-world degradations.
• Existing methods often require precise knowledge of degradation parameters, limiting practical application.

Purpose of the Study:

• To develop a comprehensive degradation model for SPI under real-world conditions.
• To create a deep-blind neural network for robust SPI image reconstruction without prior degradation knowledge.

Main Methods:

• Proposed an innovative degradation model quantifying SPI noise sources, including pattern-dependent global noise propagation and object jitter.
• Developed a deep-blind neural network trained using the comprehensive SPI degradation model.
• Implemented a method for image compensation without requiring degradation parameters.

Main Results:

• The proposed deep-blind network significantly improves SPI image resolution and fidelity.
• The method demonstrates advanced performance in real-world SPI imaging, even at ultra-low sampling rates.
• The trained network generalizes well across various combinations of degradation factors.

Conclusions:

• The developed SPI degradation model and deep-blind network offer a robust solution for high-quality image reconstruction in challenging environments.
• This approach overcomes limitations of parameter-dependent methods, enabling wider practical use of SPI.
• Potential applications span remote sensing, biomedical imaging, and surveillance.