摘要
针对电子元器件装配过程中稠密点云语义分割困难的问题,提出一种基于多邻域融合的图注意力卷积电子元器件点云分割算法。该算法通过栅格概率采样的训练策略保证点云几何结构和初始感受野;在二阶局部邻域内建立局部特征图,以图注意力卷积与池化操作逐层编码丰富语义特征;将全局特征与局部特征融合逐层上采样,完成电子元器件稠密点云的语义分割任务。实验结果表明,在电子元器件数据集上总体准确率达到96.5%,平均准确率达到93.6%,平均交并比达到85.6%,有效提高了电子元器件等稠密点云语义分割的精度。
Aiming at the difficulty of semantic segmentation of dense point clouds in the assembly process of electronic components,a graph attention convolution based multi-neighborhood fusion algorithm for point cloud segmentation of electronic components is proposed.The algorithm guarantees the point cloud geometric structure and initial receptive field through the training strategy of grid probability sampling;builds local feature maps in the second-order local neighborhood,and enriches semantic features layer by layer with graph attention convolution and pooling operations;Features and local features are fused and upsampled layer by layer to complete the semantic segmentation task of dense point clouds of electronic components.The experimental results show that the overall accuracy rate reaches 96.5%,the average accuracy rate reaches 93.6%,and the average cross-merge ratio reaches 85.6%on the electronic component data set.It effectively improves the accuracy of semantic segmentation of dense point clouds such as electronic components.
作者
沈巍
倪国林
郑金来
SHEN Wei;NI Guolin;ZHENG Jinlai(Zhongbei College,Nanjing Normal University,Danyang 212300,China;Jiangsu Gaojing Electromechanical Equipment Co.,Ltd.,Yancheng 224000,China)
出处
《组合机床与自动化加工技术》
北大核心
2025年第1期223-227,233,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
国家自然科学基金项目(51875266)。
关键词
多邻域融合
图注意力机制
电子元器件
点云分割
multi-neighborhood fusion
graph attention mechanism
electronic component
point cloud segmentation
