All-optical image classification through unknown random diffusers using a single-pixel diffractive network.

Area of Science:

• Computational imaging
• Machine vision
• Optical physics

Background:

• Classifying objects through scattering media is challenging for computational imaging and machine vision.
• Deep learning methods require significant computational resources for object classification using diffuser-distorted patterns.
• Existing approaches often rely on digital computers and deep neural networks.

Purpose of the Study:

• To present an all-optical processor for direct object classification through unknown random diffusers.
• To bypass the need for large-scale computation and deep neural networks.
• To demonstrate a system capable of classifying objects using broadband illumination and a single-pixel detector.

Main Methods:

• Developed a physical network of transmissive diffractive layers optimized using deep learning.
• The diffractive network all-optically maps object spatial information to the output light's power spectrum.
• Utilized broadband illumination and a single-pixel detector for classification.

Main Results:

• Achieved 87.74 ± 1.12% blind testing accuracy in classifying handwritten digits through unseen random diffusers.
• Experimentally validated the system using terahertz waves and a 3D-printed diffractive network for handwritten digit classification.
• Demonstrated the system's ability to process broadband light and potential for operation across the electromagnetic spectrum.

Conclusions:

• The single-pixel all-optical object classification system offers a computationally efficient alternative to deep learning methods.
• The passive diffractive layers can be adapted for various wavelengths, expanding potential applications.
• Potential applications include biomedical imaging, security, robotics, and autonomous driving.