Far-field super-resolution ghost imaging with a deep neural network constraint.

Area of Science:

• Optics and photonics
• Computational imaging
• Artificial intelligence in imaging

Background:

• Ghost imaging (GI) enables low-light image acquisition using single-pixel measurements.
• High-resolution GI typically demands extensive single-pixel sampling, limiting practical applications.
• Existing GI methods face challenges in achieving both high resolution and efficient sampling.

Purpose of the Study:

• To develop a super-resolution ghost imaging technique overcoming sampling limitations.
• To enhance image resolution beyond the diffraction limit in far-field imaging.
• To create an interpretable GI framework by incorporating physical constraints.

Main Methods:

• A hybrid deep neural network incorporating the physical model of GI image formation.
• Super-resolution reconstruction of far-field images.
• Experimental validation using a flying drone target.

Main Results:

• Reconstruction of far-field images with resolution surpassing the diffraction limit.
• Demonstrated superior performance in spatial resolution and sampling ratio compared to existing GI techniques.
• The physical model provides interpretability to the neural network output.

Conclusions:

• The proposed deep learning-based GI technique offers a new framework for high-resolution imaging.
• This method significantly improves spatial resolution and sampling efficiency in ghost imaging.
• The study paves the way for practical applications of advanced ghost imaging.