Fast and blur-resilient target classification enabled by single-pixel detection

Area of Science:

• Optical Engineering
• Computer Vision
• Signal Processing

Background:

• Conventional imaging struggles with performance degradation due to optical aberrations, defocus, and motion blur, limiting resource-constrained vision applications.
• Single-pixel detection offers a low-cost, high-speed sensing alternative but existing methods often require image reconstruction or intensive computation.
• Current single-pixel techniques primarily address motion blur and do not fully overcome image degradation challenges.

Purpose of the Study:

• To develop a fast and blur-resilient classification framework for single-pixel detection systems.
• To enable direct 'measure-to-recognize' workflows, eliminating the need for image processing or neural network training.
• To achieve high-speed optical intelligence in challenging imaging environments.

Main Methods:

• Extraction of blur-invariant features directly from single-pixel measurements.
• Utilized a system employing only seven fixed Digital Micro-mirror Device (DMD) modulation masks.
• Bypassed the conventional 'image-then-recognize' paradigm for a direct 'measure-to-recognize' approach.

Main Results:

• Demonstrated 98.96% recognition accuracy on a test dataset.
• Achieved a high update rate of 2.551 kHz.
• Validated performance under various degraded imaging conditions, including optical aberrations and motion blur.

Conclusions:

• The proposed framework offers a simple, efficient, and blur-robust solution for target classification.
• It enables high-speed optical intelligence in challenging imaging scenarios.
• Represents a novel approach to direct feature extraction and recognition from single-pixel measurements.