图残差神经网络支持下的建筑物群组模式分类

Classification of building group patterns using graph residual neural network

建筑物作为城市中的重要地物,分析其群组模式对地图综合、导航定位、市政规划等具有重要作用。建筑物群组模式分析目前主要有基于规则的方法和基于机器学习的方法两种。基于规则的方法和基于传统机器学习分类器的方法均需要大量的人工处理过程。近年来兴起的深度学习特别是图卷积神经网络前期无需人工处理,因此提高了建筑物群组模式分析的自动化程度。传统的图卷积神经网络模型在训练深层网络时易出现退化问题,提取深层特征困难。为解决此问题,本文引入了图残差神经网络模型用于建筑物群组的模式分类。首先使用道路和河流等作为约束条件,利用K-means方法对建筑物进行聚类;然后根据Bertin视觉变量计算对应的建筑物特征指标,在每个建筑物群组中以建筑物质心为节点,连接节点的最小生成树作为边,构建建筑物群组图结构;最后将得到的图结构数据输入图残差神经网络进行训练,得到规则和不规则两种建筑物群组模式。试验结果表明,该模型较好地解决了传统图卷积神经网络模型的退化问题,并取得了更高的精度。

Buildings are important features in the city,the analysis of building group patterns is of great significance in many fields such as map generalization,navigation,municipal planning and so on.Traditional methods for the recognition of building group patterns can be roughly divided into two categories:rule-based methods and machine-learning methods which require a lot of manual processing.In recent years,deep learning,especially the graph convolution neural network’s emerging that does not require prior manual processing and can improve the automation degree of the analysis of building group patterns.Traditional graph convolutional neural network model is prone to degradation when training deep networks,which makes it difficult to extract deep features.To solve this problem,a graph residual neural network(GResNet)model is proposed for the classification of building group patterns.Firstly,the roads and rivers are used as constraints,and the K-means method is used to cluster the buildings.Secondly,many indices are used to compute Bertin’s visual variables.In each building group,the centroids of the buildings are taken as nodes,and the minimum spanning tree is used to generate edges connecting nodes,after that,the graph representation for building group is constructed.Finally,the building graphs are taken as the input of the proposed GResNet model,and two building group patterns are obtained,namely,regular groups and irregular groups.Experiment results confirm that the proposed model can solve the degradation problem of the traditional graph convolutional neural network model,and obtain higher accuracy.