Multi-level efficient 3D image reconstruction model based on ViT.

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Single-photon Light Detection and Ranging (LIDAR) 3D reconstruction is hindered by high noise, low accuracy, and slow processing.
• Traditional methods are inefficient in noisy environments, and Convolutional Neural Networks (CNNs) struggle with global feature extraction.

Purpose of the Study:

• To develop a multi-level, efficient 3D image reconstruction model for single-photon LIDAR.
• To improve the quality, accuracy, and robustness of 3D reconstructions in high-noise and low-photon conditions.

Main Methods:

• A novel model based on Vision Transformer (ViT) was proposed, utilizing its self-attention mechanism for global and local feature extraction.
• Attention mechanisms were employed for feature fusion and refinement.
• Generative Adversarial Networks (GANs) were integrated to enhance reconstruction quality and robustness.

Main Results:

• The ViT-based model effectively captures both global and local features, outperforming traditional methods and CNNs in high-noise environments.
• Integration of GANs further improved reconstruction quality and robustness.
• The model demonstrated significant enhancements in real-world single-photon 3D reconstruction imaging systems under strong noise.

Conclusions:

• The proposed multi-level efficient 3D image reconstruction model based on ViT and GANs offers a significant advancement for single-photon LIDAR.
• This approach effectively addresses the challenges of noise and low photon counts, leading to superior 3D reconstruction capabilities.