基于二值神经网络的大场景点云分类

近年随着3维数据采集技术不断发展,大场景 点云数据的获取越来越方便。目前深 度学习网络框架在2维图像处理领 域越来越成熟,而大场景点云是一种3维无规则化的数据,3维卷积神经网络直接处理大场 景3维数据会存在分类精度低和计 算复杂等问题。因此为了有效解决基于深度学习的点云分类任务中存在的计算时间长和分类 精度低的问题,本文提出基于二值 神经网络的大场景点云分类方法,针对不规则的3维点云数据设计特征值计算方法,基于IR -Net二值神经网络处理输入的点云 特征图像,进一步采用Dynamic ReLU激活函数,提高神经网络的计算效率,最后得出点云分 类结果。实验结果表明,所提出 的方法在Oakland数据集上分类精度达到97.6%,在GML数据集中取得 了92.3%和97.2%的分类精度,实验结果证明Dy -ResNet 能够有效提升了点云分类的精度,减少计算的复杂度,并提高了训练效率。

Point cloud classification of large-scale scene based on binary neural network

With the continuous development of three-dimensional (3D) data acquis ition technology in recent years,the acquisition of three-dimensional large-scale scene point cloud data is becomin g more and more convenient.At present,the deep learning network framework is becoming more and more mature in the field of two-dimensional (2D) image processing,while the large- scale scene point cloud is a kind of 3D irregular data.When using 3D convolutio nal neural networks in deep learning to directly process large-scale scene 3D data,there will be problems such as low classification accuracy and high computational complexity.Therefore,to effectively solve the problems of long c omputation time and low classification accuracy in point cloud classification based on deep learning,this paper propos es a binary neural network-based classification method for large-scale scene point cloud.designing the eigenval ue calculation method for irregular 3D point cloud data,processing the input point cloud feature images based on IR-Net bin ary neural network,further adopting Dynamic ReLU activation function to improve the computational efficiency of the neural network,and finally deriving the point cloud classification results.The experimental results show that the propo sed method achieves 97.6% classification accuracy on the Oakland dataset and 92.3% and 97.2% in the GML dataset,Experime ntal results show that Dy-ResNet can effectively improve the accuracy of point cloud classification,reduce the compl exity of calculation and improve training efficiency.