Optronic convolutional neural networks of multi-layers with different functions executed in optics for image...

Area of Science:

• Optoelectronics
• Computer Vision
• Artificial Intelligence

Background:

• Deeper convolutional neural networks (CNNs) offer improved classification performance but require significant computational resources.
• Existing optical neural networks face limitations in complexity and performance.

Purpose of the Study:

• To propose and evaluate an optronic convolutional neural network (OPCNN) that integrates optical computation with electronic control to reduce computational costs.
• To demonstrate the effectiveness of OPCNN in classification tasks through simulations and experiments.

Main Methods:

• Implementing convolutional layers using a lenslet 4f system for multi-input images.
• Utilizing optical-strided convolution for downsampling layers.
• Achieving nonlinear activation by adjusting camera curves.
• Performing fully connected layers via optical dot product.
• Integrating optical computation with electronic data transmission and control.

Main Results:

• OPCNN demonstrated robust performance on classification tasks in both simulated and experimental settings.
• The proposed OPCNN achieved superior performance compared to other current optical convolutional neural networks.
• The architecture's complexity contributed to its enhanced performance.

Conclusions:

• OPCNN presents a viable hybrid approach for efficient deep learning computation.
• The scalability of OPCNN allows for the development of deeper networks for complex datasets.
• This technology offers a promising direction for high-performance, low-cost optical neural networks.