Improved PointNet with accuracy and efficiency trade-off for online detection of defects in laser processing

Structural light vision sensing is widely applied in online detecting defects of laser processing due to their anti-interference ability for laser beams. However, the existing algorithms cannot extract the characteristics of weld defects with high precision. The computing cost of large-scale point cloud data is high. The balance between them is the main challenge to achieve online detection. To improve accuracy and reduce computation costs, this study uses point cloud data with depth information and proposes a point cloud segmentation method. It is a novelty method based on PointNet framework that has been verified for laser welding defect detection. Specifically, it used the PointNet framework as the backbone. It extracted enough local features of weld defects by multi-scale feature fusion, which concatenated features from different feature extraction layers to learn enough features to improve detection accuracy. The experiments were conducted on the real dataset of welds. The results showed its competitive performance in weld bead measurement and classification segmentation, and the accuracy of this method is 97.4 %. The proposed method improved mean intersection-over-union (mIoU) by 2.1 % compared with its backbone (PointNet), indicating a better segmentation accuracy. In addition, the proposed method improved detection speed compared with PointNet++. It can reach 60 frames per second, 7.5 times faster than PointNet++ and meet the online monitoring requirements. To conclude, the new detection method based on the improved PointNet with higher accuracy and faster speed of detection has a wide application prospect thanks to its novel model.