Point Cloud Registration and Necking Defect Detection for Sheet Metal Quality Inspection

This paper presents a high-precision measurement method for sheet metal necking defect detection using dual-path feature fusion and cascade detection. To address registration challenges in feature-sparse regions, a dual-path parallel feature fusion module is proposed, which combines local geometric embeddings (LGE) with Neighbor-to-Point (N2P) attention. An adaptive multi-mechanism loss function (AMLF) is developed to enhance registration robustness in low-overlap scenarios. For defect detection, a two-stage cascade approach is employed: the Curvature-Normal Plane Analysis and Detection (CNPAD) algorithm identifies defect cores, while the Multi-Radius PCA Growing Algorithm (MRPGA) ensures complete defect extraction. A Curved Trajectory Neck Constriction Defect Measurement Algorithm (CT-NCDMA) is introduced for precise quantification of arbitrarily shaped defects. Experimental results on the 3DMatch dataset demonstrate a registration success rate of 95.4%, with relative rotation error (RRE) of 5.827° and relative translation error (RTE) of 0.152m. On industrial validation samples, the method achieves RRE of 1.169°, RTE of 0.257mm, and 100% registration success rate. Defect detection accuracy reaches 98.30% with an F1-score of 0.9592. Measurement errors for length, width, and depth are reduced to 17.39%, 27.37%, and 40.29% of baseline methods, respectively. This method provides an effective solution for automated quality inspection of stamped components.