Multi-input mutual supervision network for single-pixel computational imaging

Area of Science:

• Computational imaging
• Optical sensing
• Machine learning for imaging

Background:

• Traditional imaging systems often require complex hardware and high sampling rates.
• Single-pixel imaging offers a hardware-efficient alternative but faces challenges in reconstruction quality and sampling efficiency.
• Computational imaging methods are crucial for overcoming limitations in optical sensing.

Purpose of the Study:

• To propose a novel single-pixel computational imaging method.
• To enhance image reconstruction quality at low sampling rates.
• To enable robust imaging in challenging environments like scattering media.

Main Methods:

• Development of a multi-input mutual supervision network (MIMSN).
• Inputting 1D light intensity signals and 2D random image signals into the network.
• Utilizing mutual supervision between reconstructed signals for improved accuracy.
• Employing iterative reconstruction with generated 2D images as priors.

Main Results:

• The MIMSN learns correlations between 1D and 2D signals for information complementarity.
• Spatial information from the 2D signal reduces reconstruction uncertainty.
• High-quality image reconstruction is achieved without network pretraining.
• The method demonstrates effectiveness in low sampling rate scenarios and scattering environments.

Conclusions:

• The proposed MIMSN is a powerful tool for single-pixel computational imaging.
• The method offers a robust and efficient approach for image reconstruction, particularly in challenging optical conditions.
• This technique has significant potential for applications requiring high-quality imaging with minimal hardware and data acquisition.