一种鲁棒的多特征点云分类分割深度神经网络

doi:10.19678/j.issn.1000-3428.0060004

摘要/Abstract

摘要： 点云是一种常用的三维图形表示方式，目前点云的分类和分割是深度学习领域的研究热点。现有的深度学习方法不能有效地学习点云的局部特征，从而导致较低的分类分割精度和较差的鲁棒性。为了解决这一问题，提出一种鲁棒的多特征点云分类分割深度神经网络 RMFP-DNN。首先通过自注意力模块提取点云的局部特征，然后使用 MLP(Multilayer Perceptron, 多层感知机)提取点云的全局特征，最后通过特征融合将点云的局部特征和全局特征结合来提高分类分割的准确率和鲁棒性。采用 ModelNet40 物体分类数据集和 ShapeNet 部件分割数据集对 RMFP-DNN 进行了训练和测试。结果表明，和 PointNet、PointNet++和 DGCNN 等方法相比，RMFP-DNN 具有较好的鲁棒性和准确率。

Abstract: A point cloud is one of the common representations of 3D shapes. The classification and segmentation of point clouds is a popular research topic in the field of deep learning. While the existing methods cannot effectively learn the local features of point clouds, resulting in lower classification and segmentation accuracy and poor robustness. we propose a robust deep neural network, RMFP-DNN, for multi-feature point cloud classification and segmentation. RMFP-DNN extracts the local features between points through the self-attention module, uses the multilayer perceptron (MLP) to learn the global feature of points, and finally improves the accuracy and robustness by combining the local features and the point features through feature fusion. We train and test RMFP-DNN using the ModelNet40 dataset for object classification dataset and the ShapeNet dataset for part segmentation. Results show that RMFP-DNN has better accuracy and robustness than PointNet, PointNet++ and DGCNN.

田钰杰, 管有庆, 龚锐. 一种鲁棒的多特征点云分类分割深度神经网络[J]. 计算机工程, doi: 10.19678/j.issn.1000-3428.0060004.

TIAN Yujie , GUAN Youqing , GONG Rui. A robust deep neural network for multi-feature point cloud classification and segmentation[J]. Computer Engineering, doi: 10.19678/j.issn.1000-3428.0060004.

参考文献

[1] Qi C R, Su Hao, Mo Kaichun, et al. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation[C]. In CVPR, 2017: 77-85.
[2] Zaheer M, Kottur S, Ravanbakhsh S, et al. Deep Sets[C]. In NIPS, 2017: 3391-3401.
[3] Qi C R, Su Hao, Mo Kaichun, and Guibas L J. PointNet++: Deep hierarchical feature learning on point sets in a metric space[C]. In NIPS, 2017:5105-5114.
[4] Yue Wang, Sun Yongbin, Liu Ziwei, Sarma S E, Bronstein M M, and Solomon J M. Dynamic Graph Cnn for learning on point clouds[J].ACM Transactions on Graphics, 2019, 38(1): 146.1-146.12.
[5] D Maturana and S Scherer. VoxNet: A 3D Convolutional Neural Network for Real-Time Object Recognition[C]. In IROS, 2015:922-928.
[6] Zeng Bi and Huang Wen. An Indoor Point Cloud Segmentation Method Fusing with Multi-feature Cluster Ensemble[J]. Computer Engineering, 2018, 44(003):281-286. 曾碧，黄文. 一种融合多特征聚类集成的室内点云分割方法[J]. 计算机工程,2018, 44(003):281-286.
[7] Su Hang, Maji S, Kalogerakis E, and Learned-Miller E. Multi-view convolutional neural networks for 3D shape recognition[C]. In ICCV, 2015: 945-953.
[8] Yang Xiaowen, Li Jing, Han Xie, Han Huiyan, and Tao Qian. 3D model segmentation of convolutional neuralnetwork based on octree[J]. Computer Engineering and Design, 2020, 9(10):2663-2669. 杨晓文，李静，韩燮，韩慧妍，陶谦. 基于八叉树的卷积神经网络三维模型分割[J]. 计算机工程与设计, 2020, 9(10):2663-2669.
[9] Li Jiaxin, B M Chen, and G H Lee. So-Net: Self-organizing network for point cloud analysis[C]. In CVPR, 2018:9397-9406.
[10] Yang Jun and Dang Jisheng. Semantic segmentation of 3D point cloud based on context attention CNN[J]. Journal on Communications, 2020,7(5):195-203. 杨军，党吉圣. 基于上下文注意力 CNN 的三维点云语义分割[J]. 通信学报, 2020, 7(5):195-203.
[11] Bai Jing and Xu Haoyun. MSP-Net: multi scale point cloud classification network[J]. J Comput-Aid Desig Comput Graph, 2019, 31(11):1917-1924. 白静，徐浩钧. MSP-Net:多尺度点云分类网络[J]. 计算机辅助设计与图形学学报, 2019, 31(11): 1917-1924.
[12] Vaswani A , Shazeer N , Parmar N , et al. Attention Is All You Need[C]. In NIPS, 2017:6000-6010.
[13] Devlin J, Chang Mingwei, Lee K, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding[J]. Association for Computational Linguistics, 2019, 8(5):4171-4186.
[14] Yang Jiancheng, Zhang Qiang, and Ni Bingbing. Modeling Point Clouds with Self-Attention and Gumbel Subset Sampling[C]. In CVPR, 2019:3323-3332.
[15] Velikovi P, Cucurull G, Casanova A, et al. Graph Attention Networks[J]. In ICLR, 2018, 10(2):1710-1722.
[16] Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning[J]. In AAAI, 2017, 9(6): 4278-4284.
[17] Bahdanau D, Cho K, and Bengio Y. Neural machine translation by jointly learning to align and translate[C]. In ICLR, 2015:3156-3164.
[18] Djork-Arne, Clevert, Unterthiner T, and Hochreiter S. Fast and accurate deep network learning by exponential linear units (elus)[J]. Computer Science, 2015, 6(2):1511-1525.
[19] Zhang Ting, Qi GuoJun, Xiao Bin, and Wang Jingdong. Interleaved group convolutions[C]. In ICCV, 2017:469-480.
[20] Xie Saining, Girshick R, Piotr, Dollar, Tu Zhuowen, and He Kaiming. Aggregated residual transformations for deep neural networks[C]. In CVPR, 2017: 5987-5995.
[21] Zhang Xiangyu, Zhou Xinyu, Lin Mengxiao, and Sun Jian. ShuffleNet: An extremely efficient convolutional neural net-work for mobile devices[C]. In CVPR, 2018: 6848-6856.
[22] Wu Yuxin and He Kaiming. Group normalization[C]. In ECCV, 2018:3-19.
[23] Wu Zhirong, Song Shuran, Khosla A. 3D shapenets: A deep representation for volumetric shapes[C]. In CVPR, 2015:1912-1920.
[24] Chang A X, Funkhouser T, Guibas L. ShapeNet: An information-rich 3D model repository[C]. Computer Science, 2015:1512-1523.
[25] Huang Yuan, Da Feipeng, Tang Lin. Three-dimensional point cloud compression algorithm based on improved octree[J]. Acta Optica Sinica, 2017, 37(12):133-141.
[26] Xu Yifan, Fan Tianqi, Xu Mingye, Zeng Long, and Yu Qing. SpiderCNN: Deep learning on point sets with parameterized convolutional filters[C].In ECCV, 2018:87-102.
[27] Kingma D and Ba J. Adam: A method for Stochastic Optimization[C]. In ICLR, 2015:1412-1427.