The mr-MDA: An Invariant to Shifting, Scaling, and Rotating Variance for 3D Object Recognition Using Diffractive Deep...

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Vision

Background:

• Diffractive deep neural networks (D2NNs) excel at 2D object recognition via optical manipulation.
• Multiple-view D2NN arrays (MDAs) offer advantages for 3D object classification.
• Current optical neural networks struggle with high-speed, dynamic 3D recognition invariant to shifts, scaling, and rotation.

Purpose of the Study:

• To develop a more robust multiple-view D2NN array (mr-MDA) for efficient 3D target recognition.
• To address the limitations of existing optical neural networks in dynamic and complex environments.
• To improve the invariance of 3D object recognition to target transformations.

Main Methods:

• Proposed a novel training strategy to introduce and manage random disturbances within the optical neural network system.
• Developed and numerically verified a trained mr-MDA model.
• Focused on enhancing the stability and dynamic recognition capabilities of the D2NN architecture.

Main Results:

• The trained mr-MDA model demonstrated the effectiveness of the new training strategy.
• Numerical verification confirmed the model's ability to dynamically recognize 3D objects.
• The proposed mr-MDA achieved relatively stable 3D object recognition under simulated disturbances.

Conclusions:

• The novel training strategy effectively enhances the robustness of MDAs for 3D object recognition.
• The mr-MDA architecture provides a stable and dynamic solution for high-speed 3D target identification.
• This work advances optical neural network capabilities for complex real-world recognition tasks.