摘 要 ：目的 三维点云分类作为一项关键任务，在计算机视觉、机器人、自动驾驶等领域有着广泛的应用场景。现有的三维点云分类网络在使用边卷积进行局部特征提取时通常存在输入特征差异性小，空间结构信息提取、融合不充分等问题。针对上述问题，设计了一种结合空间结构卷积和注意力机制的点云分类网络。方法 首先，提出一种空间结构卷积，在边卷积的基础上引入邻接点之间的相对位置信息来降低输入特征相似性，而后从结构和位置两个角度分别进行特征编码，实现更具多样性的局部几何结构捕获；其次，设计了全局特征编码模块，从坐标信息中提炼全局特征信息，同时在网络中融合了注意力机制，用于关联局部与全局特征表示，有效保留了全局特征信息，实现全局特征的适应性调整。最后，将局部几何结构信息和全局位置信息进行有效的融合，获得更具代表性和差异性的特征表征。结果 设计实验在公开数据集ModelNet40上对提出的网络模型的性能进行评估，点云分类总体准确率和平均准确率分别达到93.0%和89.7%，具备良好的分类准性能和预测效率。结论 实验结果表明，空间结构卷积的使用有效增大了输入特征的多样性，位置和结构的单独编码有效提高了局部特征的表达能力。同时，提出的注意力加权方式在保留全局特征前提下实现了局部特征和全局特征的关联。综上，该网络模型有较强细粒度特征提取能力，拥有良好的分类性能。
Classification network for 3D point cloud based on spatial structure convolution and attention mechanism
wubin, liuyian, zhaojie(tianjin chengjian university)
Abstract: Objective 3D point cloud classification is a crucial task with diverse applications in computer vision, robotics, and autonomous driving. In recent years, the advancement of computing device performance has enabled researchers to apply deep learning methods to the field of 3D point cloud recognition. Deep learning-based methods currently in use for 3D point cloud classification typically divide the feature information captured by the network into two distinct parts: global features and local features. Global features refer to the overall shape and structure of the point cloud, while local features capture more detailed information about individual points. By leveraging both global and local features, these methods can achieve high accuracy in point cloud classification tasks. EdgeConv is currently the most widely used method for local feature extraction in 3D point cloud classification. This method incorporates relative position vectors into the feature encoding process to effectively capture the characteristics of local structures. However, when local structures in 3D point clouds are similar, the use of relative positions in feature encoding may result in similar features, leading to poorer classification results. Furthermore, encoding only local features may not be sufficient for achieving optimal classification results, as it is crucial to also consider the correlation between local and global features. Current methods often employ attention mechanisms to learn attention scores from global features and weight local features accordingly, which can effectively establish the correlation between local and global features. However, these methods may not fully consider the importance of global feature information and may suffer from suboptimal classification results as a result. Method To address the aforementioned challenges, this paper proposes a novel 3D point cloud classification network that leverages spatial structure convolution (SSConv) and attention mechanisms. The proposed network architecture mainly consists of two parts: a local feature encoding (LFE) module and a global feature encoding (GFE) module. The former uses SSConv to encode local features from both location and structure, while the latter learns global feature representation from raw coordinate data. Furthermore, to enable effective correlation and complementarity between the feature information, we introduce an attention mechanism that facilitates adaptive adjustment of global features through weighted operations. The LFE module is composed of two main operations: graph construction and feature extraction. The LFE module performs the K-nearest neighbor (K-NN) algorithm to identify adjacent points and construct a graph structure. SSConv is a crucial feature extraction operation that involves a multi-layer perceptron. Compared to EdgeConv, SSConv introduces a relative position vector between adjacent points. This operation effectively increases the correlation distance between raw input data, enriches local region structure information, and enhances the spatial expression ability of the extracted high-level semantic information. To capture more effective local structure features, the feature extraction is encoded separately based on structure and location. Specifically, the location encoding branch encodes the coordinate information separately to obtain richer location feature information for describing the spatial location of each point. Meanwhile, the structure encoding branch encodes the relative location vector separately to learn the structure information in the local region for describing the overall geometric structure of the local neighborhood. The global feature encoding module maps raw coordinate data to high-dimensional features, which are used as the global feature representation of the point cloud. Additionally, the module includes an attention mechanism to enhance the correlation between local and global features. Specifically, an attention weighting method is used to guide the learning of global feature information using local feature information. This operation enables the correlation and fusion between local and global feature representations while preserving the raw feature information. Result To evaluate the performance of the proposed network model, experimental validation is conducted on the publicly available ModelNet40 dataset, which consists of 9843 training models and 2468 testing models in 40 classes. The classification performance was evaluated using metrics such as overall accuracy (OA) and mean accuracy (mAcc) in the experiments. To evaluate the classification performance, the proposed model was evaluated against four pointwise methods, two convolution-based methods, two graph convolution-based methods, and four attention mechanism-based methods. The experimental results demonstrate that the proposed network exhibits a good performance on the point cloud classification task and is capable of effectively representing both local and global features. The proposed method achieves an overall accuracy of 93.0%, outperforming DGCNN by 0.1%, PointWeb by 0.7%, and PointCNN by 0.8%. Additionally, the mean accuracy of the proposed method reaches 89.7%. Furthermore, an experiment was designed to validate the efficacy of SSConv. By replacing SSConv with EdgeConv in the network architecture, the experimental results indicate a reduction in overall accuracy of 0.5% on the ModelNet40 dataset, demonstrating that SSConv is better suited for local representation than EdgeConv. Meanwhile, an experiment was designed to verify the diversity of input features of SSConv. The correlation of features was evaluated using Euclidean distance, cosine distance, and correlation distance. The results indicate that SSConv enhances the diversity among input features more effectively than EdgeConv. Furthermore, the visualization results of the intermediate layer features in the model also demonstrate that SSConv can learn more distinctive features. Conclusion The proposed network model achieves better classification results, with an overall accuracy of 93.0% and a mean accuracy of 89.7%, surpassing those of existing methods. The proposed spatially structured convolution effectively enhances the variability of input features, allowing the model to learn more diverse local feature representations of objects. The proposed global feature coding method, based on the attention mechanism, effectively adjusts features and fully extracts the relationship between local and global feature information, while preserving global features. To summarize, the proposed network model exhibits a good capability for fine-grained feature extraction and achieves good classification performance.