Fig. 1. Overall workflow chart of defect detection

Fig. 2. Experiment system for laser welding defect detection

Fig. 3. Complete sample images of typical welds. (a) Sheet butt weld;(b) thick plate butt weld; (c) bead-on-plate weld

Fig. 4. Different forms of data during data preprocessing. (a) Preprocessing of point cloud HDM data; (b) RGB images of surface defects; (c) high-density point cloud data; (d) depth images including 3D profile information of defects

Fig. 5. Structural diagram of Faster R-CNN

Fig. 6. Detection results using Faster R-CNNs based on ResNet18, ResNet50, and ResNet101

Fig. 7. Statistical results of three models. (a) Loss evolution of different models; (b) point cloud detection precisions and recall rates with different models; (c) detection precisions and recall rates of defects for point clouds and RGB images with different models; (d) detection mAPs of defects for point clouds and RGB images with different models; (e) testing time of different models

Fig. 8. Typical false negative test results of Faster R-CNN model based on ResNet50

Fig. 9. Measurement process of defect sizes. (a) RGB images; (b) point clouds; (c) depth gray images; (d) threshold segmentation; (e) locating defect areas; (f) defect feature size measurement

Fig. 10. Relative errors of defect measurement results

Table 1. Specific parameters of binocular structured light sensor

Table 2. Welding parameters

Table 3. Datasets after data augmentation

Table 4. Performance parameters of three models based on point cloud analysis

Table 5. Measurement results of defect sizes