摘要
作为一种深度神经网络结构,与卷积神经网络(CNN)相比,胶囊网络可以建立不同特征之间的空间关系,具有更好地拟合特征的能力。但是,动态路由中原有的聚类算法对初始聚类中心的选择较为敏感。针对这一问题,使用密度峰值聚类(DPC)算法对原有的聚类算法进行优化,提出DPC-CapsNet模型,以提高动态路由算法的整体性能。基于TensorFlow框架的DPCCapsNet模型的实验结果表明,结合了DPC算法的胶囊网络结构在MNIST和Fashion-MNIST数据集上均具有较快的收敛速度,以及较高的分类准确率,证明了算法的有效性,同时也说明了胶囊网络在图像分类领域的应用潜力。
As a deep neural network structure,compared with convolutional neural network(CNN),the capsule network can establish the spatial relationship between different features and has the ability to better fit the features.However,the original clustering algorithm in dynamic routing is more sensitive to the selection of the initial clustering center.To solve this problem,this paper uses the Density Peak Clustering(DPC)algorithm to optimize the original clustering algorithm,and proposes the DPC-Caps Net model to improve the overall performance of the dynamic routing algorithm.The experimental results of the DPC-Caps Net model based on the Tensor Flow framework show that the capsule network structure combined with the DPC algorithm has a faster convergence speed and a higher recognition accuracy on both the MNIST and Fashion-MNIST data sets,which proves the algorithm's performance.The effectiveness also illustrates the application potential of the capsule network in the field of image classification.
作者
林凯迪
杜洪波
朱立军
LIN Kai-di;DU Hong-bo;ZHU Li-jun(School of Science,Shenyang University of Technology,Shenyang 110870,Liaoning Province;School of Information and Computation Sciences,North Minzu University,Yinchuan 750021,Ningxia Autonomous Region)
出处
《沈阳工程学院学报（自然科学版）》
2021年第2期61-67,共7页
Journal of Shenyang Institute of Engineering:Natural Science
基金
国家自然科学基金(11861003)
宁夏自然科学基金(NZ17015)。
关键词
深度学习
图像分类
胶囊网络
动态路由
DPC算法
Deep learning
image classification
capsule network
dynamic routing
DPC algorithm
