Spectrally encoded single-pixel machine vision using diffractive networks.

Area of Science:

• Optics and Photonics
• Machine Learning
• Spectroscopy

Background:

• Traditional machine vision systems often require complex sensor arrays.
• Deep learning has shown promise in various image processing tasks.
• Diffractive optics offer unique light manipulation capabilities.

Purpose of the Study:

• To demonstrate a novel single-pixel machine vision framework using diffractive optical networks.
• To classify objects by encoding spatial information into the power spectrum of diffracted light.
• To enable image reconstruction from spectral power data.

Main Methods:

• Training deep neural networks to design diffractive layers for spatial information encoding.
• Utilizing a plasmonic nanoantenna-based detector for terahertz spectrum measurements.
• Classifying handwritten digits by analyzing spectral power at ten distinct wavelengths.
• Coupling diffractive network spectral encoding with a shallow electronic neural network for image reconstruction.

Main Results:

• Successful optical classification of handwritten digits using a single-pixel spectroscopic detector.
• Experimental validation of the single-pixel machine vision framework in the terahertz spectrum.
• Demonstration of task-specific image decompression and reconstruction from spectral data.
• Encoding of spatial information into the power spectrum of diffracted light.

Conclusions:

• The developed single-pixel machine vision framework offers a novel approach for optical object classification.
• Diffractive network-based spectral encoding can be integrated with various spectral-domain systems.
• This framework has potential for new 3D imaging and sensing modalities.