Research on the single-photon imaging algorithm based on the adaptive feature fusion residual network.

Area of Science:

• Photonics and Imaging Science
• Computer Vision and Machine Learning

Background:

• Single-photon avalanche diodes (SPADs) offer photon-level sensitivity for 3D imaging.
• Low detection efficiency and noise (ambient light, dark counts) challenge accurate depth prediction in SPADs.
• Effective signal extraction from sparse, noisy photon data requires advanced imaging algorithms.

Purpose of the Study:

• To develop an efficient single-photon imaging algorithm for highly noisy data.
• To enhance the representation of effective signals and fuse multi-level features for improved depth map generation.
• To address the challenges posed by low signal-to-noise ratios in single-photon imaging.

Main Methods:

• Proposed a novel residual network incorporating an attention learning mechanism.
• Integrated an adaptive feature fusion module for complementary fusion of shallow and deep features.
• Evaluated the network on simulated and real-world datasets under varying signal-to-background ratios.

Main Results:

• The proposed network significantly enhances effective signal representation.
• Achieved efficient fusion of shallow detail and deep semantic features.
• Demonstrated superior performance in generating reliable and accurate depth maps from noisy photon data.

Conclusions:

• The developed residual network provides a robust solution for single-photon imaging in challenging noisy conditions.
• The adaptive feature fusion and attention mechanisms are key to improving depth map accuracy.
• The network shows significant advantages for applications requiring precise depth prediction from sparse photon data.