Dynamic Attention Graph Convolution Neural Network of Point Cloud Segmentation for Defect Detection

Abstract

The traditional defect detection algorithms are suitable for regular defects, but not for workpiece defects with fuzzy features and various shapes. Point cloud segmentation is effective for 3D workpiece defect detection. In this paper, a point cloud segmentation method based on Dynamic Attention Graph convolution neural network is designed and applied to workpiece surface defect detection. Because the edge convolution module cannot capture the direction of the vector between adjacent points, we propose a new method of constructing neighbor graph based on the fusion distance. In the feature encoding stage, an attention mechanism is introduced to extract features for segmentation, and finally output the category score of each point. Experimental results show that the average prediction accuracy of the new model is improved compared to the original model.