High-resolution single-photon imaging with physics-informed deep learning.

Area of Science:

• Photonics and Imaging Science
• Artificial Intelligence in Scientific Instrumentation
• Deep Learning for Signal Processing

Background:

• High-resolution single-photon imaging is hindered by hardware complexity and noise.
• Existing methods struggle with low bit depth, low resolution, and significant noise in SPAD data.

Purpose of the Study:

• To develop a deep learning framework for super-resolution single-photon imaging using SPAD sensors.
• To enhance bit depth and overall imaging quality in challenging SPAD imaging scenarios.

Main Methods:

• Characterized SPAD electronics noise using a complex photon flow model.
• Created and utilized large-scale synthetic datasets for training deep learning models.
• Developed a deep transformer network with self-attention and gated fusion for noise removal and detail extraction.

Main Results:

• Successfully demonstrated super-resolution imaging with enhanced bit depth and quality.
• The deep transformer network effectively removed multi-source noise and recovered full-frequency details.
• Achieved state-of-the-art performance in experimental applications like microfluidic inspection and Fourier ptychography.

Conclusions:

• Deep learning integration significantly advances SPAD-based super-resolution imaging capabilities.
• The proposed method offers a robust solution for overcoming limitations in low-quality single-photon data.
• This technique holds promise for various applications requiring high-fidelity imaging from SPAD sensors.