BDRBNet: an efficient network for semantic segmentation of industrial point clouds

This paper focuses on the semantic segmentation algorithms of point clouds in industrial defect detection, studies the issues of low accuracy and insufficient feature extraction of current point-based models in industrial application scenarios, and proposes an improved model based on PointNet++ called BDRBNet. Firstly, we designed a plug-and-play downsampling feature extraction module, named BDRB (Boundary Density Residual Block). This module not only extracts features but also primarily focuses on learning the following two types of information: (1) Boundary Information: By creating a novel branch to capture the boundary features ignored by the original set abstraction levels, the ability to recognize complex-shaped workpieces is enhanced; (2) Density information: by introducing a density-based sampling method for downsampling, it strengthens the feature representation in high-density areas, enhancing the features’ representativeness and discriminative power. Moreover, to reduce the impact of class imbalance, we also employed the Poly1FocalLoss function to train the model. Through extensive experiments conducted on an autonomously collected industrial lithium battery cover welding point cloud dataset, BDRBNet outperforms PointNet++ and other advanced technologies in performance. Specifically, the Overall Accuracy (OA) and mean Intersection over Union (mIoU) have improved from 89.2% and 54.8% to 93.8% and 66.8%, with increases of 4.6% and 12%, respectively.