| 双层CNN三维刚性与非刚性模型相似性分析 |
| |
| 引用本文： | 罗 林,阮文静,庄济滔,刘万顺,秦胜伟.双层CNN三维刚性与非刚性模型相似性分析[J].教育技术导刊,2019,18(6):29-34. |
| |
| 作者姓名： | 罗 林 阮文静 庄济滔 刘万顺 秦胜伟 |
| |
| 作者单位： | 广州大学 华软软件学院，广东 广州 510990 |
| |
| 基金项目： | 广东省青年创新人才科研项目（2017KQNCX275）； 广州大学华软软件学院科学研究、教育教学项目（ky201518，ky201726）； 广东大学生科技创新项目（pdjhb0688）；广州大学华软软件学院大学生创新创业训练计划项目（DCXM2018022） |
| |
| 摘 要： | 三维模型相似性分析是计算机视觉中的重点问题，如何构建其形状特征和对比函数是难点。随着深度学习出现，通过神经网络自动提取模型特征成为研究热点。构建了双层CNN网络，首先利用热核特征函数分别构建带有颜色的刚性和非刚性训练集与测试集，其次将数据集通过双层CNN网络进行模型训练，第一层实现类别初步判定，第二层实现同一模型刚性与非刚性形变区分。为了提高分类准确度，对初步分类错误的模型引入阈值判定，将其直接排除。通过实验分析，双层CNN网络刚性与非刚性的判别准确率达到99%。实验证明，该方法在模型相似性分析上是鲁棒的，且提取的特征不受人工干扰。
|
| 关 键 词： | 三维模型 相似性分析 形状特征 CNN 热核特征 |
| 收稿时间： | 2018-09-28 |
Similarity Analysis of 3D Rigid and Non-rigid Models Based on Two-layer CNNs |
| |
| LUO Lin,RUAN Wen-jing,ZHUANG Ji-tao,LIU Wan-shun,QIN Sheng-wei.Similarity Analysis of 3D Rigid and Non-rigid Models Based on Two-layer CNNs[J].Introduction of Educational Technology,2019,18(6):29-34. |
| |
| Authors: | LUO Lin RUAN Wen-jing ZHUANG Ji-tao LIU Wan-shun QIN Sheng-wei |
| |
| Institution: | South China Institute of Software Engineering，Guangzhou University，Guangzhou 510990， China |
| |
| Abstract: | Three-dimensional （3D） model similarity analysis is a key issue in computer vision. How to construct the shape feature and analysis function of the 3D model is a challenging problem. With the advent of deep learning， it is a hot topic to extract model features automatically through the neural network. Therefore， the paper proposes a two-layer CNNs. Firstly， we construct a rigid and non-rigid training set and test set with color by the heat kernel signature （HKS）. Then a two-layer CNNs with the constructed training set is trained. The first layer of CNN implements the preliminary classification for 3D models， and the second layer achieves the distinction between the rigid and non-rigid deformation of the same model. Finally， in order to improve the accuracy of classification， the threshold is presented for excluding the 3D models directly，by which they have an error classification after the preliminary classification. The two-layer CNNs constructed in our paper has an accuracy of 99% in classification of rigid and non-rigid deformation through the experiments. The proposed method is robust in the similarity analysis of the model， and the extracted features are not subject to artificial interference. |
| |
| Keywords: | three-dimensional model similarity analysis shape feature CNN HKS |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
