基于融合图卷积的深度学习点云分类方法

Deep Learning Point Cloud Classification Method Based on Fusion Graph Convolution

针对深度学习模型PointNet仅以独立点卷积进行特征提取导致的局部特征缺乏问题,提出了一种融合空间域特征和谱域特征的图卷积深度学习模型。该模型基于空间方法和谱方法分别构造图结构,以提取不同的邻域特征,并通过融合邻域特征与独立点特征得到深层次的抽象特征,其池化层采用空间金字塔池化方法加深细粒度描述。在国际摄影测量与遥感协会提供的机载LiDAR扫描点云和多光谱航空影像上的实验结果表明,相比其他对比方法,本方法的分类效果更好,分类精度为84.3%,可实现城市场景下点云数据的有效分类。

In order to solve the problem that the deep learning model PointNet only uses independent point convolution for feature extraction,which leads to the lack of local information,a fusion graph convolution deep learning model based on spatial domain features and spectral domain features is proposed in this paper.In this model,the graph structure is constructed by spatial and spectral methods to extract different neighborhood features,the deep abstract features are obtained by fusing neighborhood features and independent point features,the spatial pyramid pooling method is used to deepen the fine-grained description in the pooling layer.Experimental results on airborne LiDAR scanning point clouds and multispectral aerial images provided by the International Photogrammetry and Remote Sensing Association show that compared with other comparison methods,the classification effect of the method is better,the classification accuracy is 84.3%,and urban scenes can be realized effective classification of point cloud data.