Anti-interference diffractive deep neural networks for multi-object recognition

Area of Science:

• Neuromorphic computing
• Optical computing
• Artificial intelligence

Background:

• Optical neural networks (ONNs) show potential for high-speed, low-power computing.
• Current ONNs struggle with multi-object recognition and interference.
• Practical applications are limited by single-object classification constraints.

Purpose of the Study:

• To develop an anti-interference diffractive deep neural network (AI D2NN) for robust multi-object recognition.
• To overcome limitations of existing ONNs in complex scenarios.
• To enable all-optical real-time target recognition.

Main Methods:

• Proposed an AI D2NN using two transmissive diffractive layers.
• Employed deep-learning strategies to differentiate targets from interference.
• Mapped target spatial information to the output light's power spectrum all-optically.
• Dispersed interference as background noise.

Main Results:

• Achieved 87.4% simulated accuracy in classifying handwritten digits with 40 interference categories.
• Demonstrated robustness against intra-class, inter-class, and dynamic interference.
• Framework is scalable to different electromagnetic wavelengths.

Conclusions:

• The AI D2NN framework effectively recognizes targets in challenging multi-object scenarios.
• This advancement is crucial for practical ONN applications in target recognition.
• Paves the way for real-time, low-power all-optical computing systems.