Global-optimal semi-supervised learning for single-pixel image-free sensing.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Single-pixel sensing provides cost-effective detection but faces challenges with large labeled datasets and complex tasks.
• Image-free sensing enhances efficiency by extracting features from compressed measurements, yet conventional methods struggle with practical limitations.

Purpose of the Study:

• To develop an image-free, semi-supervised sensing framework utilizing Generative Adversarial Networks (GANs).
• To achieve end-to-end global optimization on partially labeled datasets for improved single-pixel sensing.
• To enhance the robustness and practicality of single-pixel sensing, especially in resource-constrained scenarios.

Main Methods:

• Implementation of a semi-supervised learning framework based on GANs for image-free sensing.
• End-to-end global optimization applied to datasets with partial labeling.
• Simulation and performance evaluation using the MNIST dataset at a low sampling ratio (0.1).

Main Results:

• Achieved 94.91% sensing accuracy at a 0.1 sampling ratio with only 0.3% labeled data.
• Demonstrated superior robustness compared to conventional single-pixel sensing methods (98.49% vs. 97.36% in conventional, 94.91% vs. 83.83% in resource-constrained settings).
• Significantly reduced human effort and computational resources required for detection.

Conclusions:

• The proposed GAN-based semi-supervised framework offers a more practical and powerful detection method for single-pixel sensing.
• This approach effectively overcomes the limitations of conventional methods regarding data requirements and task complexity.
• The technique shows significant promise for robust and efficient single-pixel sensing, particularly under limited data and computational resources.