电子学报 ›› 2020, Vol. 48 ›› Issue (4): 648-653.DOI: 10.3969/j.issn.0372-2112.2020.04.005

• 学术论文 • 上一篇    下一篇

利用网格卷积特征的三维形变目标分类

史聪伟1, 赵杰煜1,2, 陈瑜1   

  1. 1. 宁波大学信息科学与工程学院, 浙江宁波 315211;
    2. 浙江省移动网络应用技术重点实验室, 浙江宁波 315211
  • 收稿日期:2019-02-25 修回日期:2019-07-21 出版日期:2020-04-25
    • 通讯作者:
    • 赵杰煜
    • 作者简介:
    • 史聪伟 男,1992年出生,浙江宁波人.宁波大学信息科学与工程学院硕士研究生,主要从事三维图形处理、模式识别相关研究.E-mail:scw_vv@126.com;陈瑜 女,1992年出生,河南安阳人.现为宁波大学信息科学与工程学院博士研究生,主要从事图形图像、深度学习等相关研究E-mail:chenyu_cycy@126.com
    • 基金资助:
    • 国家自然科学基金 (No.61571247); 浙江省自然科学基金重点项目 (No.LZ16F030001)

3D Non-rigid Object Classification with Mesh Convolution Features

SHI Cong-wei1, ZHAO Jie-yu1,2, CHEN yu1   

  1. 1. Faculty of Electrical Engineering and Computer Science, Ningbo University, Ningbo, Zhejiang 315211, China;
    2. Mobile Network Application Technology Key Laboratory of Zhejiang Province, Ningbo, Zhejiang 315211, China
  • Received:2019-02-25 Revised:2019-07-21 Online:2020-04-25 Published:2020-04-25
    • Corresponding author:
    • ZHAO Jie-yu
    • Supported by:
    • National Natural Science Foundation of China (No.61571247); Key program of National Natural Science Foundation of Zhejiang Province,  China (No.LZ16F030001)

摘要: 三维目标的形状变化给目标识别带来很大挑战,同时三维网格模型的不规则数据结构难以直接应用卷积运算提取三维目标特征.对此,本文提出了一种高效的三维形变目标的网格卷积特征表示方法,准确提取形状信息并进行分类.首先通过网格卷积运算获得形变目标中典型局部曲面形状分布,其次通过马尔科夫链对曲面形状的空间共现关系建模,从而形成三维模型的全局特征描述,最后采用支持向量机实现形变目标分类.该方法将连续多项式函数作为卷积模板,实现针对不规则数据结构的网格卷积运算,并且给出了卷积模板参数的无监督学习方法.在标准非刚性三维模型数据集SHREC10与SHREC15上的实验结果表明本文方法能有效提取三维网格模型的形状信息,分类准确率分别达到了92.88%与96.54%.

关键词: 三维形变模型, 网格卷积, 三维形状特征, 支持向量机

Abstract: 3D object recognition with shape changes is a challenging task. The irregular data structure of the mesh model prevents the operation of the conventional convolution, which brings difficulties to feature extraction of the 3D non-rigid objects. In this paper, we propose a method of mesh convolution for 3D non-rigid objects to extract shape features and use them for classification. Firstly, we obtain the distribution of typical patch shapes by the mesh convolution. Then, we model the spatial co-occurrence relationship by Markov chains to complete the global feature description. Finally, we use the support vector machine to classify the 3D objects. Our method adopts the continuous polynomial function as the convolution kernel for the irregular data structure, and learn the kernel by an unsupervised learning method. Experimental results on the standard non-rigid 3D model datasets show our method can effectively extract the features and achieve classification accuracy of 92.88% on SHREC10 and 96.54% on SHREC15,respectively.

Key words: non-rigid 3D model, mesh convolution, 3D shape feature, support vector machine

中图分类号: