Research on Semantic Segmentation of PCB Point Clouds Based on Adaptive Dynamic Graph Convolution

Point cloud data offers an accurate and intuitive means of depicting the spatial relationships between electronic components and circuit boards. Nevertheless, the irregular nature of point cloud data, coupled with the abundance of small objects and reflective areas in electronic component assembly scenes, severely restricts the application of traditional convolutional neural networks (CNNs) in point cloud semantic segmentation within such scenarios. In recent years, graph convolution networks (GCNs) have garnered increasing attention, particularly in the realm of Non-Euclidean data processing. Against this backdrop, our research proposes a point cloud segmentation method for electronic components based on Adaptive Dynamic Graph Convolution. Building upon the dynamic graph convolution network (DGCNN), our approach employs k-nearest neighbors (KNN) with multiple scales to establish the adjacency relationships among all nodes. We streamline the network architecture by removing the conversion network of DGCNN. Additionally, an adaptive feature extraction module is introduced, which seamlessly integrates an attention mechanism. This module dynamically assigns varying weights to features in different regions based on fluctuations in point cloud density, thereby enhancing the network's ability to represent regions with diverse densities. We evaluate the performance of our network against other point cloud networks using the S3DIS and PCB (Printed Circuit Board) datasets. The experimental results demonstrate that our network achieves higher point cloud segmentation accuracy than the DGCNN algorithm, especially when dealing with sparse point cloud inputs and noise. This indicates the robustness of our network. Meanwhile, it provides a preliminary assurance for the manipulator to accurately grasp electronic components, which holds significant implications for the unmanned and intelligent assembly of electronic components.