Weighted sampling-adaptive single-pixel sensing.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning

Background:

• Single-pixel sensing (SPS) with neural networks offers semantic sensing but is computationally intensive for varying sampling rates.
• Existing methods require retraining for each new sampling rate, limiting adaptability and efficiency.

Purpose of the Study:

• To develop a sampling-adaptive single-pixel sensing technique requiring only one network training for dynamic sampling rates.
• To improve the computational efficiency and adaptability of neural network-based single-pixel sensing.

Main Methods:

• Introduced a novel weighting scheme in the encoding process to characterize pattern modulation efficiencies.
• Iteratively updated modulation patterns and their corresponding weights.
• Employed an optimal pattern series with highest weights for light modulation.

Main Results:

• Achieved high classification accuracy (95.00% on MNIST, 90.20% on CCPD) at significantly reduced sampling rates (0.03 and 0.07, respectively).
• Demonstrated effective single-target and multi-target sensing with a single trained network across dynamic sampling rates.
• Validated the weighted optimization technique's efficiency and accuracy in experimental implementations.

Conclusions:

• The weighted optimization technique enables sampling-adaptive single-pixel sensing, overcoming the limitations of previous methods.
• This approach significantly enhances sensing efficiency and accuracy while reducing computational load for varied sampling rates.
• The method shows great potential for real-world applications requiring flexible and efficient optical sensing.