PointGraphNet: A Hierarchical Network for Global–Local Feature Fusion in 3D Point Cloud Classification

PointGraphNet is an end-to-end neural architecture that unifies permutation-invariant global context and dynamic local geometry for 3D point cloud understanding. The network begins with an Adaptive Feature Projection Module (AFPM) that learns a spatial transformation via a T-Net equipped with orthogonal regularization, aligning raw coordinates into a robust feature space. Next, Multi-Scale EdgeConv Blocks (MSEBs) construct and update k-NN graphs in coordinate, feature, and hybrid domains, applying shared EdgeConv operations and channel-wise max pooling to extract fine-grained neighborhood descriptors. These multi-scale features are progressively combined in a Feature Fusion & Propagation Unit (FFPU) with dense residual connections, ensuring effective gradient flow and multi-level context aggregation. A composite loss—comprising cross-entropy classification, EdgeConv smoothness regularization, and transformation orthogonality penalty—guides training. Through its hierarchical design and dynamic graph updates, PointGraphNet achieves robust fusion of global shape cues and local geometric details within a single, unified framework.