[1]李肖南,王 蕾,程海霞,等.基于 SA-PointNetVLAD 的点云分类网络[J].计算机技术与发展,2022,32(05):36-41.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 006]
　LI Xiao-nan,WANG Lei,CHENG Hai-xia,et al.Point Cloud Classification Network Based on SA-PointNetVLAD[J].,2022,32(05):36-41.[doi:10. 3969 / j. issn. 1673-629X. 2022. 05. 006]
点击复制

基于 SA-PointNetVLAD 的点云分类网络()
分享到：

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

卷:
32
期数:
2022年05期
页码:
36-41
栏目:
图形与图像
出版日期:
2022-05-10

文章信息/Info

Title:
Point Cloud Classification Network Based on SA-PointNetVLAD
文章编号:
1673-629X(2022)05-0036-06
作者:
李肖南1 ; 王 蕾1; 2 ; 程海霞1 ; 张志勇1
1. 东华理工大学 信息工学院,江西 南昌 330013;
2. 江西省核地学数据科学与系统工程技术研究中心,江西 南昌 330013
Author(s):
LI Xiao-nan1 ; WANG Lei1; 2 ; CHENG Hai-xia1 ; ZHANG Zhi-yong1
1. School of Information Engineering,East China University of Technology,Nanchang 330013,China;
2. Jiangxi Engineering Technology Research Center of Nuclear Geoscience Data Science and System,Nanchang 330013,China
关键词:
自注意力机制; 点云分类; VLAD; 关键点; 高级特征信息
Keywords:
self-attention mechanism; point cloud classification; VLAD; key points; high-level feature information
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2022. 05. 006
摘要:
三维点云数据包含着丰富的形状和比例信息,如何有效准确地对点云数据进行分类已经成为了目前计算机视觉领域的研究热点。 由于点云在非欧氏 空间中的不规则稀疏结构,并且现有的基于深度学习的三维点云分类模型中缺乏对各个点的局部信息和全局信息的有效利用, 从而导致较低的分类精度。 为了解决上述问题, 提出了一种基于 SA -PointNetVLAD 的点云分类模型框架。 该网络采用逐点特征提取和汇集操作来解决无序的点云问题,通过自注意力机制来计算每个点与其他所有点之间的关联,充分挖掘点云的局部区域细粒度特征以及全局信息,采用 KNN 邻近算法感知点云的局部形状结构,通过 VLAD 层将每个点的低层几何描述符与视觉单词相关联来间接描述高层特征信息。 此外,该网络还引入一个有效的关键点描述符帮助识别整体几何图形。 经过实验得出 SA-PointNetVLAD 模型在公开的 ModelNet40 数据集上仅使用 512 个点就可以达到 90. 9% 的精度,显著高于相同条件下的其他方法。
Abstract:
3D point cloud data contains rich shape and scale information. How to classify the point cloud data effectively and accurately is a research hotspot in the field of computer vision. Due to the irregular and sparse structure of point cloud in non -Euclidian space,the existing 3D point? ? cloud classification model based on deep learning lacks the effective use of the local and global information of each point,which leads to low classification accuracy. To solve the above problems,a point cloud classification model framework based on SA-Point Net VLAD is proposed.? ?The network uses point-by-point feature extraction and aggregation operations to solve the disordered point cloud problem. By computing the association between each point and all other points through the self - attention mechanism,the local fine-grained features and global information of the point cloud are fully explored,the local shape structure of the point cloud is perceived by the KNN proximity algorithm,and the high-level feature information is indirectly described by associating the low-level geometry descriptor of each point with several visual words through the VLAD layer. In addition, the network introduces an effective key point descriptor to help identify the overall geometry. The experiments show that the SA-Point Net VLAD model can achieve 90. 9% accuracy using only 512 points on a public Model Net 40 dataset,significantly higher than other methods under the same conditions.

相似文献/References:

[1]郑树挺,徐菲菲.基于改进 Self-Attention 的股价趋势预测[J].计算机技术与发展,2021,31(03):33.[doi:10. 3969 / j. issn. 1673-629X. 2021. 03. 006]
　ZHENG Shu-ting,XU Fei-fei.Research on Stock Price Trend Prediction Based on Self-Attention Model[J].,2021,31(05):33.[doi:10. 3969 / j. issn. 1673-629X. 2021. 03. 006]
[2]秦昊宇,葛 瑶,张力波,等.基于自注意力机制的视频超分辨率重建[J].计算机技术与发展,2022,32(08):42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 08. 007]
　QIN Hao-yu,GE Yao,ZHANG Li-bo,et al.Video Super-resolution Reconstruction Based on Self Attention Mechanism[J].,2022,32(05):42.[doi:10. 3969 / j. issn. 1673-629X. 2022. 08. 007]
[3]黄 莉,何美玲*.基于 U-Net 改进模型的多模态脑肿瘤分割方法[J].计算机技术与发展,2022,32(11):58.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 009]
　HUANG Li,HE Mei-ling*.Multi-model Brain Tumor Segmentation Method Based on Improved U-Net Model[J].,2022,32(05):58.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 009]
[4]李 飞,陈 瑞,童 莹,等.基于增强特征和注意力机制的视频表情识别[J].计算机技术与发展,2022,32(11):183.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 027]
　LI Fei,CHEN Rui,TONG Ying,et al.Video Facial Expression Recognition Based on ECNN-SA[J].,2022,32(05):183.[doi:10. 3969 / j. issn. 1673-629X. 2022. 11. 027]
[5]李丰翼,刘万里,杨晓辉,等.基于多头自注意力机制与 CNN 的文本分类模型[J].计算机技术与发展,2022,32(S1):18.[doi:10. 3969 / j. issn. 1673-629X. 2022. S1. 004]
　LI Feng-yi,LIU Wan-li,YANG Xiao-hui,et al.Text Classification Model Based on Multi-headed Self-attention Mechanism and CNN[J].,2022,32(05):18.[doi:10. 3969 / j. issn. 1673-629X. 2022. S1. 004]
[6]赵建强,朱万彤,陈 诚.基于多重卷积神经网络模型的命名实体识别[J].计算机技术与发展,2023,33(01):187.[doi:10. 3969 / j. issn. 1673-629X. 2023. 01. 028]
　ZHAO Jian-qiang,ZHU Wan-tong,CHEN Cheng.Named Entity Recognition Based on Duplex Convolution Neural Network Model[J].,2023,33(05):187.[doi:10. 3969 / j. issn. 1673-629X. 2023. 01. 028]
[7]陈 奔,谢晓尧,刘 嵩.基于动态图卷积的混合注意力点云特征学习网络[J].计算机技术与发展,2023,33(10):157.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 024]
　CHEN Ben,XIE Xiao-yao,LIU Song.3D Point Cloud Feature Learning Network Based on Hybrid Attention and Dynamic Graph Convolution[J].,2023,33(05):157.[doi:10. 3969 / j. issn. 1673-629X. 2023. 10. 024]
[8]黄汉琴,顾进广,符海东.融合依存句法和实体信息的临床时间关系抽取[J].计算机技术与发展,2024,34(01):128.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 019]
　HUANG Han-qin,GU Jin-guang,FU Hai-dong.Extraction of Clinical Temporal Relation Fusing Dependency Syntax and Entity Information[J].,2024,34(05):128.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 019]
[9]刘 洋,黎茂锋,黄 俊,等.面向方面级情感分析的双通道图卷积网络[J].计算机技术与发展,2024,34(03):49.[doi:10. 3969 / j. issn. 1673-629X. 2024. 03. 008]
　LIU Yang,LI Mao-feng,HUANG Jun,et al.Aspect-based Sentiment Analysis via Dual-channel Graph Convolutional Network[J].,2024,34(05):49.[doi:10. 3969 / j. issn. 1673-629X. 2024. 03. 008]

更新日期/Last Update: 2022-05-10