Dynamic single-photon 3D imaging with a sparsity-based neural network.

Area of Science:

• Photonics
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning shows promise for single-photon LiDAR, offering high photon efficiency and image quality.
• Current deep learning methods for LiDAR face challenges with high memory usage and slow inference, limiting their use in dynamic, long-range imaging on constrained devices.

Purpose of the Study:

• To propose an efficient neural network architecture for single-photon LiDAR that addresses memory and speed limitations.
• To enable high-resolution, dynamic imaging with resource-constrained devices.

Main Methods:

• Developed an efficient neural network architecture exploiting data sparsity by skipping inactive sites with no photon counts.
• Implemented a one-shot processing approach for high-resolution data frames.
• Validated the method using public real-world datasets and a home-built system.

Main Results:

• Achieved over 90% acceleration in computation speed without sacrificing reconstruction quality.
• Demonstrated that the method's speed is independent of detection distance.
• Showcased outstanding dynamic imaging capabilities, orders of magnitude faster than competing methods.

Conclusions:

• The proposed efficient neural network architecture significantly reduces storage and computation overhead for single-photon LiDAR.
• This method is compatible with resource-constrained devices and excels in dynamic imaging applications.
• The approach offers a substantial speed improvement over existing methods without requiring data pruning.