整合集成预测约束与错误预测熵最大化的MLS点云分类方法  

Integrating ensemble prediction constraints and error prediction entropy maximization for MLS point cloud classification

作　　者：雷相达 管海燕 董震[2] LEI Xiangda;GUAN Haiyan;DONG Zhen(School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing,Wuhan University,Wuhan 430079,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]武汉大学测绘遥感信息工程国家重点实验室,武汉430079

出　　处：《遥感学报》2025年第1期329-340,共12页NATIONAL REMOTE SENSING BULLETIN

基　　金：国家自然科学基金(编号:41971414);江苏省研究生科研创新项目(编号:KYCX23_1361)。

摘　　要：许多深度学习点云分类方法通过增加点云特征聚合模块,增强点云特征的表达能力,但该类方法往往会带来训练参数增加以及模型过拟合的问题。针对该问题,本文提出了一个整合集成预测约束与错误预测熵最大化的深度学习方法用于移动激光扫描MLS(Mobile Laser Scanning)点云分类。方法通过集成预测约束分支以及错误预测熵最大化分支可以在不增加训练参数的情况下,增强基线网络的点云特征表达,提高模型泛化能力。其中集成预测约束分支首先通过记录点云在训练过程中的预测值,生成集成预测值,然后采用一致性约束增强模型的点云特征表达。错误预测熵最大化分支鼓励模型对错误预测点进行熵值最大化,增加该点的不确定性,提高模型的泛化能力。所提方法在多个公开MLS点云数据集上进行验证,结果表明所提方法可以在不增加训练参数的情况下,提高基线方法的分类性能。与对比方法相比,所提方法在Toronto3D、WHU-MLS、Paris数据集上获得了最优的平均交并比(83.68%、65.85%、44.19%),表明了方法的有效性。Mobile Laser Scanning(MLS)systems are widely used in various fields owing to their ability of rapidly acquiring highprecision and high-density 3D point cloud data,particularly in the acquisition of urban spatial information.Given that urban MLS point clouds exhibit complex scenes,large data volumes,and uneven spatial distribution,accurate classification of large-scale urban point clouds presents significant challenges.Currently,many deep learning point cloud classification methods enhance feature representation of point clouds by adding a feature aggregation module.Nonetheless,this approach frequently results in increased training parameters and model overfitting.We propose an MLS point cloud classification method integrating resemble prediction constraints and error prediction entropy maximization.The proposed method can enhance the point cloud feature representation of the baseline network and improve the generalization ability of the model without increasing the training parameters.Our method consists of three main components:a basic supervision branch,an ensemble prediction constraint branch,and an error prediction entropy maximization branch.Specifically,we first employ RandLA-Net as the backbone network to obtain point cloud classification features.Then,a basic supervised branch calculates the weighted cross-entropy loss on the basis of true labels and predicts probability distributions and category weights to provide a basic fully supervised signal for model training.For the ensemble prediction constraint branch,we first generate ensemble predictions by recording the predicted values during the point cloud training process.Because the input to RandLA-Net is a random subpoint cloud,the ensemble predictions can be integrated for predictions not only at different stages but also at different relative positions.Thus,the ensemble prediction is highly robust to the current prediction.Afterward,we apply a consistency constraint to minimize the difference between the two predictions to improve the point cloud feat

关 键 词：遥感 MLS点云分类 深度学习 集成预测约束 错误预测熵最大化 

分 类 号：P237[天文地球—摄影测量与遥感] P2[天文地球—测绘科学与技术]