Infrared object classification with a hybrid optical convolution neural network.

Area of Science:

• Computer Vision
• Machine Learning
• Optical Engineering

Background:

• Machine vision systems face challenges balancing high-resolution imaging needs with embedded system constraints like power and bandwidth.
• Current systems struggle to meet the demands for both detailed image data and efficient processing in resource-limited environments.

Purpose of the Study:

• To design and validate a multi-pixel hybrid optical convolutional neural network (CNN) system for high-speed infrared object detection.
• To address the trade-off between high-resolution image acquisition and the limitations of embedded systems.

Main Methods:

• A hybrid optical CNN system was developed, utilizing a high-speed digital micro-mirror device (DMD) to replicate the front convolution layer.
• The system displays initial kernel layers at a resolution exceeding the subsequent detector, with further convolutions performed in software.
• Performance was validated using an infrared vehicle dataset through simulation and hardware testing with toy vehicles.

Main Results:

• The hybrid optical CNN system demonstrated feasibility for high-speed infrared object detection.
• Simulation results using an infrared vehicle dataset confirmed the system's effectiveness.
• Hardware tests on toy vehicles showcased the practical application of the designed system.

Conclusions:

• The proposed hybrid optical CNN system offers a viable solution for high-speed infrared object detection in embedded systems.
• This approach effectively mitigates the constraints of power, bandwidth, and storage while maintaining detection performance.
• The system represents a significant advancement in machine vision for resource-constrained applications.