PSDNet: A Bio-Inspired Perception-Based Method for Multi-Scale Semantic Segmentation of Non-Uniform Point Clouds of Precast Components

Three-dimensional laser-scanned point clouds of precast components are often affected by non-uniform density, large object-scale variation, and construction-site noise. Inspired by the dual-pathway perception of eagle vision, this study proposes PSDNet, an end-to-end semantic segmentation network for precast-component point clouds. An Eagle-Focused Density Intensification module (EFDI) enhances panel-edge details in sparsely sampled regions, while a Multi-Scale Global Fusion Module (MSGM) combines a one-time-built spatial index with global dual-pooling attention to distinguish objects ranging from centimeter-scale embedded components to meter-scale fixtures. Experiments were conducted on a dataset collected from a precast-component project in Riyadh, Saudi Arabia. PSDNet achieved a mean intersection-over-union of 73.6%, panel extraction accuracy of 90.9%, and embedded-component identification rate of 91.8%. Ablation experiments confirmed the complementary contributions of EFDI and MSGM. Comparisons with PointNet++, PointCNN, RandLA-Net, and KPConv showed consistently superior segmentation under non-uniform point densities and large-scale variations. The results demonstrate the potential of PSDNet for digital inspection of precast components.