胶囊网络综述

打开文本图片集

关键词：胶囊网络；卷积神经网络；姿态；向量胶囊网络；矩阵胶囊网络；优化策略中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2025）10-001-2881-12doi：10.19734/j.issn.1001-3695.2025.03.0066

Overview of capsule networks

Zhu Xiaojuana，Li Weijun a，b† ，Ma Xinyua， Xiong Zhangyoua，Yang Guolianga （a.Schoolofutee&Enin，Ktoofges&aellgentrocessnofateEcs sion，NorthMinzuUniversity，Yinchuan75Oo21，China）

Abstract：Capsulenetworksefectivelyovercome thedatadependencyofconvolutionalneuralnetworksandtheir limitations in modeling object position，pose，and spatial relationships through their unique network structure and dynamic routing algorithm，demonstrating significant potentialandadaptabilty.Tosystematicallreviewtheresearchprogressofcapsulenetwork's architecture，thisrticlediscusedfundamentaltoriesndoptimiationtrategies，experimentalcomparisons，andaation domains.Firstly，itintroducedthebasicarchitecturesandcorealgorithmsofvectorcapsulenetworksandmatrix capsule networks，andaalyzedvariousoptimizationstrategiesalong withtheiradvantagesanddisadvantages.Next，experimentalcomparisons evaluated the performance ofcapsule networksand their improved models.Furthermore，thispaper explored applicationsofcapsulenetworksinbectdtection，heathcare，andtrasportation.Finalytoutlinesfutureresearchdectiosnd development trends to guide further advancements in the field.

Key words：capsulenetwork ；convolutional neural network；pose;vectorcapsule network ；matrix capsule network；optimization strategies

0 引言

卷积神经网络（convolutional neural network，CNN）[1，2]在计算机领域取得了突破性进展，在计算机视觉、机器学习等多个领域得到了广泛的应用[3]，如图像处理[4]、自然语言处理[5]等。（剩余36948字）