WS-CCNet: A central-channel network for weakly supervised point cloud segmentation of workpiece surface defects

Point cloud-based surface defect segmentation is essential for high-precision automated inspection and processing in industrial applications. Most existing point cloud segmentation methods rely on dense point-level annotations, leading to substantial labeling costs in real-world deployment. Despite growing interest in weakly supervised learning, effective semantic propagation from sparse annotations remains challenging. We propose WS-CCNet, a weakly supervised central-channel network for accurate segmentation of protruding defects on cast workpiece surfaces. In WS-CCNet, the central-channel dual attention (CCDA) mechanism couples center-guided spatial aggregation with deviation-aware channel enhancement, enabling more discriminative feature propagation under sparse supervision. In addition, a geometry-aware position encoding (GAPE) module encodes multi-level geometric information into the attention mechanism to improve sensitivity to subtle protrusion defects. A dropout-based regularization strategy is further introduced as a lightweight auxiliary constraint to encourage prediction consistency under stochastic perturbations. Experimental results on the workpiece dataset show that WS-CCNet significantly outperforms existing approaches for surface defect segmentation under sparse supervision, while competitive results on ShapeNetPart, S3DIS, and SemanticKITTI further suggest its generalization potential.