三维点云深度学习分类模型学习策略分析——以机载多光谱LiDAR点云为例  

Learning strategies analysis of 3D point cloud deep learning classification model:a case study of airborne multi-spectral LiDAR point cloud

作　　者：梁家安 陈科 季铮[2] 管海燕 LIANG Jiaan;CHEN Ke;JI Zheng;GUAN Haiyan(School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430072,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,南京210044 [2]武汉大学遥感信息工程学院,武汉430072

出　　处：《测绘工程》2023年第6期68-75,共8页Engineering of Surveying and Mapping

基　　金：国家自然科学基金资助项目(41971414)。

摘　　要：近年来,随着三维计算机视觉的发展,三维点云深度学习方法得到越来越多学者的关注。然而,目前大多数三维点云深度学习方法仅在标准数据集上进行精度和性能评估,这些方法在设计过程中通常会根据特定数据集设定特定的学习策略以达到最佳点云分类精度,从而影响模型的泛化能力。在三维点云深度学习方法的遥感应用中,往往会出现诸如复杂的网络模型方法并不一定取得更好的数据处理精度,以及原始网络模型学习策略并不一定获得最优结果等问题。对此,为了探究学习策略对三维点云深度学习方法在实际遥感应用的影响,文中以机载多光谱LiDAR点云分类应用为例,以深度学习经典模型PointNet++为主,在分析当前学习策略的基础上构建一套较为通用的深度学习策略,以提高点云分类精度的稳定性和鲁棒性。机载多光谱LiDAR点云分类实验表明,学习策略对于点云分类精度影响不容忽视,学习策略的调整可以有效地提高模型对海量三维点云分类能力。In recent years,with the development of three-dimensional(3D)computer vision,LiDAR deep learning method has received more and more scholars’attention.However,at present,most 3D point cloud deep learning methods only evaluate the accuracy and performance on standard data sets.These methods usually use specific learning strategies according to feature data sets in the design process to achieve the optimal point cloud classification accuracies,thus affecting the generalization ability of the models.Specifically,in the practical remote sensing applications of LiDAR deep learning methods,there are often some problems,such as the accuracies obtained by more complex network methods are not necessarily higher,and the results obtained by the learning strategies of original networks are not the optimal.To explore the influence of learning strategies on 3D point cloud learning method in the actual remote sensing applications,this paper adopts airborne multispectral LiDAR data as example,and builds a general learning strategy based on the PointNet++.The deep learning strategy is used to improve the stability and robustness of point cloud classification tasks for remote sensing applications.Airborne multi-spectral LiDAR classification experiments showed that the impact of the learning strategy on point cloud classification accuracy cannot be ignored,and the adjustment of learning strategy can effectively improve the model’s ability to classify massive 3D point clouds.

关 键 词：三维图像处理 多光谱LiDAR 深度学习 点云分类 数据增强 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]