摘要
针对传统点云分类网络提取特征较为单一,分类精度较低的问题,提出一种多层次多尺度点云分类网络MLMS-Net。首先使用预处理算法将原始点云分割为小样本,以得到批处理输入,提高训练效率;然后使用K近邻算法和边缘特征向量分别提取点云的低层次结构特征和边缘特征,通过设置不同邻域值,有效获取上下文信息,通过点内和点间多层次表达获得局部细粒度描述;接着对两种低层次特征分别构建卷积神经网络,随着网络层次的加深,特征抽象程度越来越高,区分程度也随之增加,从而有效提高准确性;最后利用后处理模块融合深层特征,完成点云分类任务。使用Vaihingen数据集对MLMS-Net网络进行测试,其分类精度相较单层次网络提高了0.6%~15.9%。
A point cloud classification network with multi-level and multi-scale(MLMS-Net)is proposed to improve the problem that the traditional point cloud classification network extracts single-level features and teh classification accuracy is low.First,apreprocessing algorithm is used to segment the original point cloud into small samples to obtain mini-batch and so as to improve training efficiency;Then the K-nearest neighbor algorithm and an edge feature vector are used to extract the low-level structure features and edge features of the point cloud,and the context information is effectively obtained through setting different neighborhood values.Local fine-grained descriptions are obtained through multi-level expression within and between points.Then convolutional neural networks are constructed for two low-level features separately.With the deepening of the network level,the degree of feature abstraction level is higher and higher,and the degree of distinction increases,so as to effectively improve the accuracy;Finally,the post-processing module is used to fuse the deep features,and the point cloud classification task is completed.The Vaihingen data set is used to test the MLMS-Net network,its classification accuracy is improved by 0.6%to 15.9%compared with that of the single-level network.
作者
薛豆豆
程英蕾
文沛
余旺盛
秦先祥
XUE Doudou;CHENG Yinglei;WEN Pei;YU Wangsheng;QIN Xianxiang(Information and Navigation College,Air Force Engineering University,Xi’an 710077,China;The 93575 Unit,Chengde,Hebei 067000,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2020年第12期70-78,共9页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金青年科学基金资助项目(61703423)
国家自然科学基金资助项目(41601436)。
关键词
点云分类
卷积神经网络
边缘特征
局部细粒度
point cloud classification
convolutional neural network
edge features
local fine-grained
