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

Single-photon LIDAR faces challenges in high-quality 3D reconstruction due to high noise levels, low accuracy, and long inference times. Traditional methods, which rely on statistical data to obtain parameter information, are inefficient in high-noise environments. Although convolutional neural networks (CNNs)-based deep learning methods can improve 3D reconstruction quality compared to traditional methods, they struggle to effectively capture global features and long-range dependencies. To address these issues, this paper proposes a multi-level efficient 3D image reconstruction model based on vision transformer (ViT). This model leverages the self-attention mechanism of ViT to capture both global and local features and utilizes attention mechanisms to fuse and refine the extracted features. By introducing generative adversarial ngenerative adversarial networks (GANs), the reconstruction quality and robustness of the model in high noise and low photon environments are further improved. Furthermore, the proposed 3D reconstruction network has been applied in real-world imaging systems, significantly enhancing the imaging capabilities of single-photon 3D reconstruction under strong noise conditions.