用于三维点云识别的双模块图卷积网络

Dual-module graph convolutional network for 3D point cloud recognition

针对三维点云识别的无序性且缺乏拓扑信息的问题。设计了一种双模块图卷积神经网络,能够同时学习点云的多尺度特征和残差特征,以提高三维(3D)点云的识别精度。在第一个模块中,利用基于修正余弦相似度的图卷积,以及不同的特征维度,提取点云的多尺度特征。在第二个模块中,利用多层网络间的跨层连接,以及固定尺度邻域图,提取点云残差特征。将两种特征融合,作为多层感知机的输入,经过池化层和全连接层网络,对点云进行分类。在公开数据集ModelNet40上的实验结果,证明了该算法的有效性。

Aiming at the problem of disorder and lack of topological information of 3 D point cloud recognition, a dual module graph convolutional neural network is devised to improve 3 D point cloud recognition accuracy, by learning both multi-scale features and residual features of point cloud.In the first module, the multi-scale features are extracted via graph convolution based on adjusted cosine similarity and different feature dimensions.In the second module, the residual features of point cloud are extracted by the skip connections of multi-layer networks and the fixed-scale neighborhood graphs.Then, two kinds of features are fused as the input of multilayer perceptron(MLP).Finally, the point cloud is classified via a pooling layer and fully connected layers.The experimental results on the public available dataset ModelNet40 demonstrate the effectiveness of the proposed approach.