| 基于语义信息补偿全局特征的物体点云分类分割 |
| |
| 引用本文： | 林森,赵振禹,任晓奎,陶志勇.基于语义信息补偿全局特征的物体点云分类分割[J].红外与激光工程,2022,51(8):20210702-1-20210702-12. |
| |
| 作者姓名： | 林森 赵振禹 任晓奎 陶志勇 |
| |
| 作者单位： | 1.沈阳理工大学 自动化与电气工程学院，辽宁 沈阳 110159 |
| |
| 基金项目： | 国家重点研发计划(2018 YFB1403303) |
| |
| 摘 要： | 3D点云数据处理在物体分割、医学图像分割和虚拟现实等领域起到了重要作用。然而现有3D点云学习网络全局特征提取范围小，难以描述局部高级语义信息，进而导致点云特征表述不完整。针对这些问题，提出一种基于语义信息补偿全局特征的物体点云分类分割网络。首先，将输入的点云数据对齐到规范空间，进行数据的输入转换预处理。然后，利用扩张边缘卷积模块提取转换后数据的每一层特征，并叠加生成全局特征。而在局部特征提取时，利用提取到的低级语义信息来描述高级语义信息和有效几何特征，用于补偿全局特征中遗漏的点云特征。最后，融合全局特征和局部高级语义信息得到点云的整体特征。实验结果表明，文中方法在分类和分割性能上优于目前经典和新颖的算法。
|
| 关 键 词： | 语义信息 3D模型分类、分割 特征提取 深度学习 |
| 收稿时间： | 2022-01-20 |
Object point cloud classification and segmentation based on semantic information compensating global features |
| |
| Affiliation: | 1.School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China2.School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China |
| |
| Abstract: | 3D point cloud data processing has played an essential role in object segmentation, medical image segmentation, and virtual reality. However, the existing 3D point cloud learning network has a small global feature extraction range and cannot obtain local high-level semantic information, which leads to incomplete point cloud feature representation. Aiming at these problems, a classification, and segmentation network of object point cloud based on semantic information compensating global features was proposed. Firstly, align the input point cloud data to the specification space, and perform the preprocessing of the input conversion of the data. Then, the expanded edge convolution module was used to extract the features of each layer of the converted data and superimpose them to generate global features. In the local feature extraction, the extracted low-level semantic information was used to describe the high-level semantic features and effective geometric information, which was used to compensate for the missing point cloud features in the global features. Finally, the global feature and local high-level semantic information were combined to obtain the overall feature of the point cloud. The experimental results show that the method in this paper is superior to the current classic and novel algorithms in classification and segmentation performance. |
| |
| Keywords: | |
|
| 点击此处可从《红外与激光工程》浏览原始摘要信息 |
|
点击此处可从《红外与激光工程》下载全文 |
|
