[1]范文豪,吴晓富,张索非.动态路由胶囊网络的可视化研究[J].计算机技术与发展,2019,29(08):71-75.[doi:10. 3969 / j. issn. 1673-629X. 2019. 08. 014]
 FAN Wen-hao,WU Xiao-fu,ZHANG Suo-fei.Research on Visualization of Capsule Network with Dynamic Routing[J].,2019,29(08):71-75.[doi:10. 3969 / j. issn. 1673-629X. 2019. 08. 014]
点击复制

动态路由胶囊网络的可视化研究()
分享到:

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
29
期数:
2019年08期
页码:
71-75
栏目:
智能、算法、系统工程
出版日期:
2019-08-10

文章信息/Info

Title:
Research on Visualization of Capsule Network with Dynamic Routing
文章编号:
1673-629X(2019)08-0071-05
作者:
范文豪1 ;?吴晓富1 ;?张索非2
1. 南京邮电大学 通信与信息工程学院,江苏 南京 210003; 2. 南京邮电大学 物联网学院,江苏 南京 210003
Author(s):
FAN Wen-hao1 ;?WU Xiao-fu1 ;?ZHANG Suo-fei2
1. School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China; 2. School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China
关键词:
胶囊网络;?矢量化;?空间信息;?可视化摇
Keywords:
capsule network;?vectorization;?spatial information;?visualization
分类号:
TP391
DOI:
10. 3969 / j. issn. 1673-629X. 2019. 08. 014
摘要:
最近提出的胶囊网络是继卷积网络(convolutional neural networks,CNN)之后的另一创新结构。 相比于 CNN 特征的弱空间关联性,胶囊网络的矢量化特征则被认为能很好地表达特征之间的空间关联。 然而,胶囊网络的特征应如何理解还缺乏严格的理论以及实验论证。 对此,文中试图从特征可视化的角度来验证 CNN 中特征空间关联性的强弱,并探讨胶囊网络提取到的特征中是否具有空间关联。 通过训练不同特征维数的胶囊网络,探究出改变特征维数对胶囊网络产生的影响。 实验结果表明,相比于 CNN 特征空间的弱关联性,胶囊网络的矢量化特征呈强相关性,确实包含了所提取特征的姿态、形变等空间相关信息,并且当胶囊网络特征维数降低时提取到的空间信息会减少,使得胶囊网络复原出图像的与输入的原图像差距增大。
Abstract:
Recently proposed capsule networks provide an alternative to convolutional neural networks (CNN). Compared with the weak spatial correlation of CNN features,the vector feature of capsule network is considered to be an effective way to express the spatial correlation between features. However,how to understand the characteristics of the capsule network still lacks strict theoretical and experimental supports. Therefore,we attempt to verify the strength of feature space association in CNN and explore whether the features extracted from the capsule network have spatial connections from the perspective of feature visualization. The influence of changing the dimension of features on the capsule network is explored by training the capsule network with different dimensions of features. The experiment shows that compared with the weak correlation of CNN feature space,the vectorization of capsule network is strongly correlated,including spatial correlation information such as attitude and deformation of extracted features. When the dimension of features is reduced,the spatial information extracted is reduced,so that the gap between the image restored by the capsule network and the input original image is increased.

相似文献/References:

[1]赵重庆 张凯龙 周兴社 姚远.新型保真花样格式优化与多样式缝制控制[J].计算机技术与发展,2010,(02):137.
 ZHAO Chong-qing,ZHANG Kai-long,ZHOU Xing-she,et al.An Improved High- Fidelity Pattern Format and Multi- Style Sewing Control[J].,2010,(08):137.
[2]段友祥,李宁宁,孙歧峰,等.一种油藏地质剖面图矢量化方法[J].计算机技术与发展,2018,28(08):170.[doi:10.3969/ j. issn.1673-629X.2018.08.036]
 DUAN You-xiang,LI Ning-ning,SUN Qi-feng,et al.A Method of Vectorization of Oil Reservoir Geological Profile[J].,2018,28(08):170.[doi:10.3969/ j. issn.1673-629X.2018.08.036]
[3]于 灏,杨建鸣,王小刚.基于改进果蝇算法的工程图纸分割方法研究[J].计算机技术与发展,2018,28(10):124.[doi:10.3969/ j. issn.1673-629X.2018.10.026]
 YU Hao,YANG Jian-ming,WANG Xiao-gang.Research on Engineering Drawing Segmentation Method Based on Improved Fruit Fly Optimization Algorithm[J].,2018,28(08):124.[doi:10.3969/ j. issn.1673-629X.2018.10.026]

更新日期/Last Update: 2019-08-10