Robust and efficient single-pixel image classification with nonlinear optics.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• High-resolution image processing demands efficient and accurate classification methods.
• Existing techniques may struggle with noisy or occluded data, limiting their practical applications.

Purpose of the Study:

• To develop a hybrid image classifier for fast and robust processing of high-resolution images.
• To improve classification accuracy and performance in challenging imaging conditions.

Main Methods:

• A hybrid approach combining feature-sensitive image upconversion, single-pixel photodetection, and deep learning.
• Utilizing partial Fourier transform for extracting signature features in both spatial and Fourier domains.
• Testing on Modified National Institute of Standards and Technology (NIST) handwritten digit datasets.

Main Results:

• Achieved a classification accuracy of 99.23%, a significant improvement from 81.25% with Fourier-domain processing alone.
• Demonstrated 83% accuracy for highly contaminated images with a signal-to-noise ratio as low as -17 dB.
• Validated performance through simulation and testing on benchmark datasets.

Conclusions:

• The proposed hybrid classifier offers superior accuracy and robustness for high-resolution image recognition.
• The method shows potential for applications in lidar data processing, occluded target identification, and atmospheric monitoring.
• This approach addresses limitations of current methods in handling degraded image quality.