Physics-driven deep learning for high-fidelity photon-detection ghost imaging.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning Applications

Background:

• Single-photon detection offers high sensitivity for imaging but faces challenges in resolution through scattering media due to low light, high noise, and detector time jitter.
• Existing methods struggle to achieve high spatial and depth resolution in complex scattering environments.

Purpose of the Study:

• To develop a physics-driven, learning-based photon-detection ghost imaging method overcoming limitations in scattering media.
• To enhance both spatial and depth resolution for improved imaging performance.

Main Methods:

• Co-design of a computational ghost imaging system and a neural network for integrated imaging and reconstruction.
• Utilizing fringe patterns to encode object depth information into an image cube.
• Employing a specialized depth fusion network with attention mechanisms for feature extraction.

Main Results:

• Achieved super-resolution reconstruction at 256x256 pixels.
• Demonstrated superior imaging performance across various scenarios, surpassing physical resolution limitations.
• Validated the method's effectiveness in challenging scattering conditions.

Conclusions:

• The proposed method offers a compact and cost-effective alternative for photon-detection imaging in scattering media.
• The integrated approach successfully enhances resolution and image quality.
• This technique holds significant potential for advancing ghost imaging applications.