StructuredLight-GNN: A graph neural network for end-to-end structured light 3D point cloud processing

To address the core challenges of complex noise, insufficient feature utilization, and poor task collaboration in structured light 3D scanning, this paper presents a graph neural network, named StructuredLight-GNN. By integrating confidence-aware mechanisms, attention mechanisms, and multi-scale feature extraction modules, the network constructs an end-to-end framework that enables collaborative processing across the entire “denoising - optimization - detection” pipeline for structured light point clouds. Key innovations include confidence-fused adaptive graph construction, multi-branch heterogeneous feature extraction, multi-task joint optimization, and hierarchical multi-scale processing. The network effectively handles complex noise patterns in structured light scans, such as phase unwrapping errors, ambient light-induced outliers, Gray code decoding failures, and blocky missing points caused by highly reflective surfaces. It also accurately identifies imaged objects even in the presence of obstacles and interference, demonstrating the significant potential of graph neural networks in structured light 3D scanning.