结合双注意力和加权动态图卷积的三维点云分类与分割

3D point cloud classification and segmentation based on dual attention and weighted dynamic graph convolution

针对基于深度学习的点云分类分割网络在局部上下文信息提取和近邻点特征表达上的不足,以及最大池化容易丢失次优信息的问题,提出结合双注意力和加权动态图卷积网络的点云分类分割算法。首先,加权动态图卷积利用加权K近邻算法构建鲁棒的局部结构,并引入强化边卷积模块对点特征加权以得到强化后边特征。然后,通道注意力构造通道相关性并释放各通道潜力,再利用空间注意力感知三维点云的空间结构,以增强局部语义特征的表达,并提取有效上下文信息与深层语义特征。最后,采用TopK池化添加次优特征。实验结果表明,该算法在ModelNet40分类数据集上总体分类精度达到93.36%,在ShapeNet Part部件分割数据集上平均交并比达到85.96%,能够有效提取上下文信息和增强近邻点特征表达,表明了算法的有效性。

In response to the deficiencies in local context information extraction and neighboring point feature expression in deep learning-based point cloud classification and segmentation networks,as well as the problem of maxpooling leading to the loss of suboptimal information,a point cloud classification and segmentation algorithm that combines dual attention and weighted dynamic graph convolutional networks was proposed.Firstly,the weighted dynamic graph convolution used a weighted k-nearest neighbor algorithm to construct a robust local structure and introduced an enhanced edge convolution module to apply weights to point features,thereby obtaining enhanced edge features.Then,channel attention was used to construct channel correlations and unleash the potential of each channel,followed by spatial attention to perceive the spatial structure of 3D point clouds,enhancing the expression of local semantic features and extracting effective contextual and deep semantic information.Finally,TopK pooling was employed to add suboptimal features.Experimental results show that the algorithm achieves an overall classification accuracy of 93.36%on the ModelNet40 classification dataset and an average intersection over union of 85.96%on the ShapeNet Part segmentation dataset,effectively extracting contextual information and enhanced neighboring point feature expression,demonstrating the effectiveness of the algorithm.