基于深度学习的抗噪声点云识别网络设计  

Design of anti-noise point cloud recognition network based on deep learning

作　　者：张光玺 汤汶 万韬阮[3] 薛涛[1] ZHANG Guangxi;TANG Wen;WAN Taoruan;XUE Tao(School of Computer Science,Xi’an Polytechnic University,Xi’an 710048,China;Faculty of Science and Technolgy,Bournemouth University,Poole BH125BB,United Kingdom;Faculty of Engineering and Informatics,University of Bradford,Bradford BD71DP,United Kingdom)

机构地区：[1]西安工程大学计算机科学学院,陕西西安710048 [2]伯恩茅斯大学科学与技术学院,英国伯恩茅斯BH125BB [3]布拉德福德大学工程与信息学院,英国布拉德福德BD71DP

出　　处：《纺织高校基础科学学报》2020年第3期113-120,共8页Basic Sciences Journal of Textile Universities

基　　金：陕西省科技厅自然科学基金(2016JZ026);陕西省科技厅国际科技合作与交流计划(2016KW-043)。

摘　　要：为了提高点云识别网络的抗噪声能力,降低神经网络在空间模型运算中对处理器的压力,设计一款轻量且具备抗噪声能力的点云识别网络。新的网络通过引入点云库技术,在多层感知机输入数据前添加了随机采样模块和近邻统计高斯滤噪模块,有效滤除复杂点云场景中的离群点。通过优化多层感知模块与全连接模块层次结构,减少网络冗余参数。实验证明:在模型识别准确率维持在84.2%的同时,相较于7种同类型网络,本网络对数据中的随机噪声具有较强的鲁棒性,并具有更快的识别速度。In order to improve the anti-noise ability of the point cloud recognition network and reduce the pressure of the neural network on the processor in the spatial model operation,a lightweight and anti-noise point cloud recognition network was designed.By introduceing the Point Cloud Library,adding a random downsampling module and a StatisticalOutlierRemoval filter module before the input data of the multi-layer perceptron,the outliers in complex point cloud scenese were effectively filtered out.By optimizing the hierarchical structure of multi-layer perception modules and fully connected module,the redundant parameters of network are reduced.The experimental results show that compared with seven kinds of networks of the same type,this network has stronger robustness to the random noise in the data and has faster recognition speed while the accuracy rate of model recognition is maintained at 84.2%.

关 键 词：点云识别 抗噪声 轻量型 点云库 随机下采样 高斯统计滤噪 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]