Single-Photon Imaging in Complex Scenarios via Physics-Informed Deep Neural Networks.

Area of Science:

• Photonics and Computational Imaging
• Deep Learning for Scientific Applications

Background:

• Single-photon imaging captures 3D structure using sensitive sensors but struggles in complex environments.
• Traditional methods degrade, and deep learning approaches lack flexibility and generalization in challenging scenarios.

Purpose of the Study:

• To develop a robust framework for single-photon imaging in complex environments.
• To enhance 3D reconstruction accuracy and generalization capabilities.

Main Methods:

• Proposed a physics-informed deep neural network (PIDNN) framework integrating imaging physics for unsupervised learning.
• Tailored U-Net skip connections for multi-scale spatiotemporal priors to improve photon efficiency.
• Incorporated volume rendering and a dual-branch structure for multi-depth and fog scenarios.

Main Results:

• Achieved robust performance in low signal-to-background ratio (SBR) and large fields of view with photon-efficient imaging.
• Demonstrated lower root mean-squared error compared to traditional methods.
• Exhibited superior generalization and reconstruction quality over supervised methods in multi-depth and fog conditions.

Conclusions:

• The PIDNN framework offers a flexible and scalable solution for complex single-photon imaging challenges.
• Validated through simulations and experiments, the method shows exceptional reconstruction performance and adaptability.
• Successfully addresses limitations in traditional and supervised deep learning for 3D scene reconstruction.