Real-time target recognition with all-optical neural networks for ghost imaging

Area of Science:

• Optics
• Computational Imaging
• Artificial Intelligence

Background:

• Computational ghost imaging relies on random speckle patterns for image reconstruction.
• Traditional electronic hardware limits the modulation capabilities for speckle pattern generation.
• All-optical neural networks offer a novel approach to overcome these modulation limitations.

Purpose of the Study:

• To propose and demonstrate a real-time target recognition system for ghost imaging using an all-optical diffraction deep neural network.
• To address the limitations of traditional electronic hardware in modulating speckle patterns.
• To evaluate the system's performance and feasibility in practical applications.

Main Methods:

• Utilizing a trained all-optical neural network for pure phase modulation of visible light.
• Implementing target recognition by detecting maximum light intensity signals at various positions.
• Optimizing system parameters through simulation, including network layers and unit area.
• Materializing the neural network using 3D printing technology.

Main Results:

• Achieved a target recognition accuracy of 91.73% after parameter optimization.
• Demonstrated successful real-time target recognition at a low sampling rate of 1.25%.
• Verified the system's feasibility and noise resistance through experimental validation.

Conclusions:

• The proposed all-optical deep neural network system effectively performs real-time target recognition in ghost imaging.
• The system overcomes the modulation limitations of conventional electronic hardware.
• The 3D-printed neural network demonstrates practical applicability and robustness in real-world scenarios.