«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年03期

页码:

49-55

栏目:

媒体计算

出版日期:

2025-03-10

文章信息/Info

Title:

Local Feature Enhancement Network for Point Cloud Semantic Segmentation

文章编号:

1673-629X(2025)03-0049-07

作者:

柴玉晶; 梁坤豪; 杨历省; 宫卫光

枣庄学院 光电工程学院,山东 枣庄 277160

Author(s):

CHAI Yu-jing; LIANG Kun-hao; YANG Li-sheng; GONG Wei-guang

School of Photovoltaic Engineering,Zaozhuang College,Zaozhuang 277160,China

关键词:

图像处理; 点云语义分割; 局部特征增强; 反密度函数; 深度学习; 特征向量

Keywords:

image processing; point cloud semantic segmentation; local feature enhancement; inverse density function; deep learning; feature vector

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0340

摘要:

相比二维图像数据,三维点云数据普遍具有无序性、稀疏性和密度不均匀性,同时点与点之间的相关性以及结构信息的获取也是很大的挑战。 正是由于点云的这些特性,对于特征信息不显著的物体的语义分割,一直是点云处理领域的一大难题。 为此,提出了一种用于点云语义分割的局部特征增强网络,该算法设计了一种局部特征聚合模块用于增强特征信息。 该模块将点云的相对点位置与其对应点特征串联在一起,从而获得增强的特征向量,实现了特征信息的增强。通过该模块使网络能够有效地学习复杂的局部结构信息,增强对局部几何信息的处理能力。 此外,该算法在提取特征时采用反密度函数对稀疏区域的点赋予较大的权重,对稠密区域的点赋予较小的权重,从而有效消除了点云的密度不均匀性造成的影响。 实验结果表明,该算法在斯坦福大规模三维室内空间数据集的平均交并比达到了 67. 4% ,整体准确率达到了 88. 4% ,相比 DGCNN 分别提升了 11. 3% 和 4. 3% ,对特征信息不显著的目标物的分割效果提升显著。

Abstract:

Compared to two-dimensional image data,3D point cloud data generally has disorder,sparsity,and uneven density. At the same time,the correlation between points and the acquisition of structural information are also great challenges. It is precisely because of these characteristics of point clouds that semantic segmentation of objects with insignificant feature information has always been a major challenge in the field of point cloud processing. To this end,a local feature enhancement network for point cloud semantic segmentation is proposed. The algorithm designs a local feature aggregation module to enhance feature information. This module concatenates the relative point positions of the point cloud with its corresponding point features to obtain an enhanced feature vector, achieving the enhancement of feature information. Through this module,the network can effectively learn complex local structural information and enhance its ability to process local geometric information. In addition,the proposed algorithm uses an inverse density function to assign larger weights to points in sparse regions and smaller weights to points in dense regions when extracting features,effectively eliminating the influence of uneven density of point clouds. The experimental results show that the proposed algorithm achieves an average intersection-union ratio of 67. 4% and an overall accuracy of 88. 4% on the Stanford large-scale 3D indoor spatial dataset,which is 11.3% and 4. 3% higher than DGCNN respectively. It significantly improves the segmentation effect of objects with insignificant feature information.

常用功能

更新日期/Last Update: 2025-03-10