LOEN: Lensless opto-electronic neural network empowered machine vision

Area of Science:

• Optoelectronics
• Machine Vision
• Artificial Intelligence

Background:

• Machine vision systems face challenges with high power consumption and large data volumes.
• Existing opto-electronic hybrid neural networks often require complex structures and coherent light sources, limiting their use in natural lighting.

Purpose of the Study:

• To propose a novel lensless opto-electronic neural network architecture for efficient machine vision applications.
• To overcome the limitations of existing systems in terms of power consumption, data handling, and environmental adaptability.

Main Methods:

• Developed a lensless opto-electronic neural network architecture.
• Optimized a passive optical mask using task-oriented neural network design.
• Implemented optical convolution calculations within the lensless architecture.

Main Results:

• Achieved 97.21% accuracy in handwritten digit classification using a multiple-kernel mask.
• Demonstrated effective optical encryption for privacy-preserving face recognition with equivalent accuracy to non-encrypted methods.
• Improved recognition accuracy by over 6% compared to random MLS patterns.

Conclusions:

• The proposed lensless opto-electronic neural network offers a compact and computationally efficient solution for machine vision.
• The architecture is suitable for natural lighting environments and enhances privacy in applications like face recognition.