基于图卷积的计算机辅助设计模型分类

计算机辅助设计(CAD)模型是一种带有顶点信息和网格信息的三维数据,三维模型数据存储方式常见的有点云、体素、网格模型等是典型的非欧氏空间数据。为了改进现有方法利用深度学习训练CAD模型的分类时,常有丢失局部信息或局部信息提取不足的情况。针对这种非欧氏空间的CAD数据,提出了一个结合CAD数据本身特点的基于图卷积的分类模型。首先通过图卷积网络(GCN)计算顶点的邻接矩阵和顶点的度矩阵。针对CAD模型的特点提出了不同于K近邻(KNN)的方法,直接根据CAD模型面片信息构建计算所需的邻接矩阵。其次,图卷积网络可以聚合邻近顶点的信息,设计通过拼接两层图卷积网络来提取不同尺度的局部特征。结果表明:在ModelNet40 CAD模型数据集上,若采用CAD模型面片信息建图的方法,本文方法为91.2%。而采用KNN建图的方法虽然比PointNet++模型低1%的精确度,比KD-NET模型低0.9%的精确度,但参数量要比PointNet++减少0.54 MB,比KD-NET减少6.54 MB。可见本文模型结合了CAD模型的特点和图卷积聚合邻接顶点提取局部信息的优势,使得分类的精确度相比PointNet++提高0.6%,用更少的模型参数量得到了更高的分类精确度。

CAD Model Classification based on Graph Convolution

Objective Previous deep learning methods training CAD model classification, often lost local information or local information extraction is insufficient. CAD model is a kind of 3D data with vertex information and grid information, The storage mode of 3d data, such as point cloud, voxel and mesh model, is typical non-euclidian spatial data, In view of this kind of CAD data in non-euclidean space, a classification model based on graph convolution is proposed based on the characteristics of CAD data.Method Firstly, the calculation of graph convolution network (GCN) requires the adjacency matrix of vertices and degree matrix of vertices. This paper proposes a method different from the k-nearest neighbor (KNN) method according to the characteristics of CAD model, and constructs the adjacency matrix needed for calculation directly based on the surface information of CAD model. Secondly, graph convolutional network can aggregate information of adjacent vertices. We designed a two-layer graph convolutional network to extract local features of different scales Result On ModelNet40 CAD model data set, if using the method of CAD model surface information to establish graph, the graph convolution classification method proposed in this paper is 91.2%. Although the accuracy of KNN establish graph method is 1% lower than that of PointNet++ model and 0.9% lower than that of KD-NET model, the number of parameters is 0.54MB lower than PointNet++ and 6.54MB lower than KD-NET.Conclusion In this paper, the classification model based on graph convolution is proposed, which combines the features of CAD model and the advantage of graph convolution aggregation of adjacent vertices to extract local information, so that the classification accuracy is increased by 0.6% compared with PointNet++, and higher classification accuracy is obtained by using fewer model parameters