Single-pixel neural network object classification of sub-Nyquist ghost imaging.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computational Imaging

Background:

• Traditional ghost imaging requires numerous measurements to reconstruct images, limiting its speed and applicability.
• Sub-Nyquist sampling presents challenges for accurate object recognition due to insufficient data.
• Neural networks offer powerful tools for pattern recognition and data analysis.

Purpose of the Study:

• To develop a single-pixel neural network for object classification within a sub-Nyquist ghost imaging framework.
• To demonstrate high classification accuracy without full image reconstruction.
• To enhance processing speed and robustness against noise.

Main Methods:

• Utilizing a single-pixel camera and random illumination patterns in a ghost imaging setup.
• Implementing a neural network for direct classification from bucket measurements.
• Employing a parallel computing scheme for accelerated data processing.
• Evaluating performance under sub-Nyquist conditions and environmental noise.

Main Results:

• Achieved 94.23% classification accuracy with only 16 measurements, significantly below the Nyquist limit.
• Demonstrated superior performance for both binary and grayscale images compared to existing sub-Nyquist methods.
• Showcased robustness against environmental noise and turbulence.
• Significantly reduced object acquisition time through parallel computing.

Conclusions:

• The proposed single-pixel neural network method enables efficient and accurate object classification in sub-Nyquist ghost imaging.
• This approach overcomes limitations of traditional ghost imaging and sub-Nyquist sampling for recognition tasks.
• The method's speed, accuracy, and robustness suggest potential applications in remote sensing and defense.