结合局部增强与反向残差的点云语义分割网络  

Combining local augmentation and reverse residual for point cloud semantic segmentation network

作　　者：温智成 王蕾[1] 冯锦梁 叶森辉 WEN Zhi-cheng;WANG Lei;FENG Jin-liang;YE Sen-hui(School of Information Engineering,East China University of Technology,Nanchang 330013,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013

出　　处：《激光与红外》2023年第7期1131-1136,共6页Laser & Infrared

基　　金：江西省核地学数据科学与系统工程技术研究中心基金项目(No.JELRGBDT202202);江西省放射性地学大数据技术工程实验室开放基金项目(No.JELRGBDT202103)资助。

摘　　要：激光点云是3D传感器的输出,且对它的语义分割任务是理解真实世界的基础。基于图卷积的点云分割网络在许多场景下都展现了优异的性能。然而,现有的图卷积方法存在部分问题:点云局部表示的能力未得到加强,忽略了全局几何信息,并且聚合操作只保留局部最大响应值信息,而次最大值信息丢失。为了处理这些问题,本文提出GRes-Net网络。利用局部几何加强(Local Geometry Augment,LGA)模块,使网络对Z轴具有旋转不变性,以便加强点云局部信息表示;采用全局几何特征(Global Geometry Feature,GGF)模块,计算局部与全局的球体体积比,将其与坐标特征X进行连接,使全局几何信息特征得以保留;通过多个对称聚合操作将局部信息多方面地保留;网络中每层都使用残差操作,将上一层信息传递到下一层,以及利用反向残差模块(Reversed Residual MLP,RevResMLP)挖掘更深层次的语义信息。本文在S3DIS数据集上进行语义场景分割实验,验证网络分割的性能。实验结果表明该方法在分割精度上达到61%,相比于基准网络DGCNN提高14%,有效地提高了模型性能。The laser point cloud is the output of a 3D sensor,and its semantic segmentation task is fundamental to understanding the real world.The point cloud segmentation network based on graph convolution shows excellent performance in many scenarios.However,existing graph convolution methods are partially problematic:the local representation of the point cloud is not augmented,global geometric information is ignored,and the aggregation operation only keeps the local maximum response value information,while the sub-maximum information is lost.In order to address these problems,the GRes-Net network is proposed in this paper.The Local Geometry Augment(LGA)module is used to make the networkrotationally invariant to the Z-axis,so as to strengthen the local information representation of the point cloud.The Global Geometry Feature(GGF)module is used to calculate the volume ratio between local and global spheres,and connect it with coordinate feature X,so that the global geometry information feature can be retained.Multiple symmetric aggregation operations are used to preserve the local information in many ways.Residual operations are used in each layer of the network to transfer information from one layer to the next as well as to mine deeper semantic information using the Reversed Residual MLP(RevResMLP)module.In this paper,semantic scene segmentation is conducted on S3DIS data set to verify the performance of network segmentation.The experimental results show that the segmentation accuracy of this method is up to 61%,which is 14%higher than that of the benchmark network DGCNN,and the performance of the model is effectively improved.

关 键 词：点云语义分割 动态图卷积 局部几何加强 全局几何特征 反向残差 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN249[自动化与计算机技术—计算机科学与技术]