摘要
胶囊网络是一种新型深度神经网络,采用向量表达图像特征信息,并通过引入动态路由算法解决了卷积神经网络的两个主要问题:1)无法对图像的部分与整体关系进行学习和表达;2)池化操作导致图像特征信息严重丢失。然而,CapsNet需要学习图像的所有特征,当图像背景较复杂时,其存在提取图像特征信息不足、训练参数量大和训练效率低等问题。为此,首先设计了一种轻量级的图像特征提取器RA模块,用于更快、更完整地提取图像特征信息;其次,设计了两种不同深度的轻量化分支来提升网络的训练效率;最后,设计了新的压缩函数hc-squash来确保网络能够获取更多有用信息,并提出了多分支RA胶囊网络。通过在MNIST,Fashion-MNIST,affNIST和CIFAR-10这4个图像分类数据集中的应用,证实了多分支RA胶囊网络在多项性能指标上优于CapsNet和MLCN,并针对所提网络设计了改进方案,以优化分类性能。
Capsule Network is a new type of deep neural network that uses vectors to express information of image feature and overcomes two major problems of convolutional neural networks by introducing dynamic routing algorithms.First,convolutional neural networks cannot learn and express the part-whole relationship of images.Second,pooling operations lead to serious loss of image feature information.However,CapsNet needs to learn all the features of the image,and when the image background is complex,it has the problems of insufficient information of extracted image features,large number of training parameters and low training efficiency.To this end,firstly,a lightweight image feature extractor RA module is designed to extract image feature information faster and more completely.Secondly,two different depths of lightweight branches are designed to improve the training efficiency of the network.Finally,a new compression function hc-squash is designed to ensure that the network can acquire more useful information,and a multi-branch RA(Resnet Attention)capsule network is proposed.Through the application in the four image classification datasets of MNIST,Fashion-MNIST,affNIST and CIFAR-10,it is confirmed that the multi-branch RA capsule network outperforms CapsNet and MLCN in several performance metrics,and an improvement scheme is designed for the proposed network to achieve optimised classification performance.
作者
武霖
孙静宇
WU Lin;SUN Jing-yu(College of Software,Taiyuan University of Technology,Taiyuan 030024,China)
出处
《计算机科学》
CSCD
北大核心
2022年第6期224-230,共7页
Computer Science
关键词
胶囊网络
RA模块
深度学习
压缩函数
注意力机制
Capsule network
Resnet attention module
Deep learning
Squash function
Attention mechanism
