摘要
多边形网格作为一种复杂的数据结构为三维物体提供了有效的形状近似表示,但由于网格数据的复杂性和不规则性,卷积神经网络很难直接应用到三维网格数据处理中.因此,提出一种基于胶囊网络的深度学习方法对三维网格数据进行有效分类.首先设计一种能够直接在网格表面进行计算的多项式卷积模板,提取三维网格模型的高阶参数特征.同时,为解决传统卷积神经网络大量池化层的引入导致的特征丢失问题,针对输入网格模型尺寸大小不统一问题,改进了胶囊网络姿态参数学习方法,通过共享姿态矩阵权值,减少模型参数量,进一步提高了三维网格模型的训练效率.实验在SHREC15数据集上与传统方法和最新的两种方法进行比较,相比于最新方法MeshNet和MeshCNN,在原始测试集上的平均识别准确率提高了3.4和2.1个百分点,且融合特征后平均准确率达到93.8%.经实验验证,在训练时间较短的前提下,该方法也能取得相当的识别效果.本文所提出的三维网格分类方法,综合了图形学与深度学习方法的优点,有效提高了三维网格模型的分类效果.
As a complex data structure,polygon mesh provides an effective shape approximate representation for three-dimensional objects.However,due to the complexity and irregularity of mesh data,convolutional neural networks are difficult to directly apply to 3D mesh data processing.Considering this problem,we proposed a deep learning method based on capsule network to effectively classify 3D mesh data.First,design a polynomial convolution template that can directly calculate on the surface of the mesh and extract the height of the 3D mesh model.At the same time,in order to solve the problem of feature loss caused by the introduction of a large number of pooling layers of traditional convolutional neural networks,the capsule network is introduced.Aiming at the problem of inconsistent size of the input grid model,the capsule network attitude parameter learning method is improved,and the weight of the pose matrix reduces the amount of model parameters and further improves the training efficiency of the three-dimensional mesh model.The experiment is compared with the traditional method and the latest two methods on the SHREC15 data set.Compared with the latest methods MeshNet and MeshCNN,the average recognition accuracy of the original test set has increased by 3.4 and 2.1 percentage points,and the average accuracy after fusion of features reaches 93.8%.Under the premise of shortening the training time,after experimental verification,this method can also achieve a considerable recognition effect.The proposed three-dimensional grid classification method combines the advantages of graphics and deep learning methods,and effectively improves the classification effect of the three-dimensional grid model.
作者
郑阳
赵杰煜
陈瑜
唐晨
俞书世
ZHENG Yang;ZHAO Jieyu;CHEN Yu;TANG Chen;YU Shushi(Faculty of Electrical Engineering and Computer Science,Ningbo University,Ningbo 315211,China;Mobile Network Application Technology Key Laboratory of Zhejiang Province,Ningbo 315211,China)
出处
《厦门大学学报（自然科学版）》
CAS
CSCD
北大核心
2021年第4期711-719,共9页
Journal of Xiamen University：Natural Science
基金
国家自然科学基金(62071260)。
关键词
胶囊网络
池化
三维模型识别
特征提取
capsule network
pooling
three-dimensional model recognition
feature extraction
