Improved pointnet-based binocular structured light detection for ceramic ball defects

Abstract

To address low accuracy in defect detection of point clouds from binocular structured light 3D reconstruction caused by high reflectivity of the surface of silicon nitride ceramic balls, this study proposes a method integrating PointNetLK point cloud registration and PointNet++ defect recognition networks. Initially, PointNet segments and removes reflective regions from the point clouds. Subsequently, an enhanced PointNetLK network performs high-precision binocular point cloud registration with missing region compensation, demonstrating two orders of magnitude improvement in registration accuracy over the conventional Coherent Point Drift (CPD) + Iterative Closest Point (ICP) methods. Finally, the compensated complete point clouds are processed by an enhanced PointNet++ network incorporating a Multi-Scale Grouping (MSG) strategy for defect segmentation, effectively identifying two primary defect types (pits and scratches) with an average mIoU of 0.8565. Ablation studies confirm the critical contributions of the Set Abstraction (SA) module and MSG strategy. This approach significantly mitigates hyper-reflection interference, achieving high-precision, robust, and non-destructive quantification of ceramic ball surface defects.