High-frame-rate reconfigurable diffractive neural network based on superpixels.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Science

Background:

• Reconfigurable diffractive neural networks typically use liquid-crystal spatial light modulators and digital micromirror devices, leading to complex optical systems and slow computation.
• Existing methods face challenges in optical system alignment and limited processing speeds, hindering real-time applications.

Purpose of the Study:

• To propose a simplified superpixel diffractive neural network architecture.
• To enhance computational speed in diffractive neural networks.
• To demonstrate the efficacy of the proposed method in image and action recognition tasks.

Main Methods:

• A novel superpixel diffractive neural network design utilizing solely a digital micromirror device.
• Controlling neuron bias and connections via the digital micromirror device for simplified optical setup.
• Experimental validation using digit classification and action recognition datasets.

Main Results:

• Achieved a computational speed of 326 Hz per neural layer.
• Digit classification accuracy reached 82.6%.
• Perfect 100% accuracy was obtained for action recognition tasks.

Conclusions:

• The superpixel diffractive neural network significantly simplifies optical system design.
• The proposed method substantially increases computational speed compared to existing implementations.
• This advancement holds promise for real-time optical information processing applications.