Single-pixel perception method based on sorting pattern and meta-classifier correction.

Area of Science:

• Optics
• Computer Vision
• Machine Learning

Background:

• Fixed encoding patterns in compressed sensing limit information utilization and affect classification accuracy.
• Effectiveness of compressed sensing varies across encoding patterns and datasets, impacting feature information and perception system efficiency.

Purpose of the Study:

• To propose a sequential optimization approach for encoding patterns to enhance information acquisition and classification accuracy.
• To address limitations in current encoding patterns for improved feature extraction and perception system performance.

Main Methods:

• A sequential optimization scheme for encoding patterns was developed.
• Encoding patterns were combined with category representative images for intra-class feature sorting.
• Inter-class feature differences were identified to reorder encoding patterns for enhanced feature extraction.
• A meta-classifier was used for direct classification after feature extraction with reordered patterns.

Main Results:

• The optimized encoding pattern sequence achieved high classification performance with only 20 samples.
• Simulation results showed 88.20% and 80.00% accuracy on MNIST and FASHION MNIST datasets, respectively.
• Practical optical experiments confirmed classification accuracies of 87.50% and 80.00% for MNIST and FASHION MNIST with 20 samples.

Conclusions:

• The proposed encoding pattern sequence optimization significantly improves classification accuracy under low sampling rates.
• This method enhances the efficiency and effectiveness of single-pixel perception classification systems.
• The findings provide a foundation for future advancements in single-pixel perception classification.