DHGCNN: Dynamic Hypergraph Convolutional Network for Industrial 3D Defect Segmentation

Existing industrial 3D defect segmentation models, which rely on point-based or graph-based approaches to capture pairwise relationships, fail to effectively represent the inherent complex higher-order geometric structures in the data. To overcome this limitation, we develop a Dynamic Hypergraph Convolutional Neural Network (DHGCNN), within which two synergistic hypergraph components are jointly employed to achieve robust and exhaustive high-order feature representation. Specifically, the proposed HyPConv module constructs global hyperedges to capture long-range dependencies and structural correlations across the entire point cloud, while the HyperEdge-Conv module focuses on refining local geometric relationships within each hyperedge by dynamically aggregating multi-point contextual information. These modules jointly enable DHGCNN to achieve more expressive representation learning, further enhanced by a residual learning strategy and a channel attention mechanism that improve feature propagation and discriminative focus. Extensive experiments on both real-world (Real3D-AD) and synthetic (Anomaly-ShapeNet) datasets demonstrate that DHGCNN significantly outperforms existing baselines in both segmentation accuracy and IoU metrics.