多尺度自路由胶囊网络的构建方法  

Construction Method of Multi-Scale Self-Routing Capsule Network

作　　者：张贤坤[1] 陶健伟 董梅 袁菁 ZHANG Xiankun;TAO Jianwei;DONG Mei;YUAN Jing(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2022年第3期59-66,共8页Journal of Tianjin University of Science & Technology

基　　金：天津市自然科学基金项目(19JCYBJC1530);天津市普通本科高校教学改革和教学质量研究计划项目(B201005706)。

摘　　要：胶囊网络(capsule network,CapsNet)是近些年新提出的一种网络模型.相比于传统卷积神经网络,胶囊网络能有效编码不同特征的空间位置关系,但是其特征提取模块难以捕捉不同尺度的特征信息.为了进一步提升CapsNet的性能,提出了一种多尺度自路由胶囊网络(multi-scale self-routing capsule network,MSSR-CapsNet).首先利用多分支结构改进CapsNet胶囊的生成过程,增加模型多尺度特征提取能力;然后提出了一种胶囊注意力机制,选择性筛选出重要的低级特征;最后在SVHN、CIFAR-10和CIFAR-100数据集上对MSSR-CapsNet进行实验,分类准确率分别为97.46%、91.22%和63.84%.此外,实验结果也验证了改进方法的高效性,有助于提升模型的性能.Capsule network(CapsNet)is a new network model proposed in recent years.Compared with the traditional convolutional neural network structure,CapsNet can effectively encode the spatial relationship of different features,but its feature extraction module is difficult to capture feature information of different scale.In order to further improve the performance of CapsNet,a multi-scale self-routing capsule network(MSSR-CapsNet)is proposed in this article.First,the multibranch structure is used to improve the capsule generation process of CapsNet and increase the multi-scale feature extraction ability of model;then a capsule attention mechanism is proposed to selectively filter out important low-level features.Finally,on SVHN,CIFAR-10 and CIFAR-100,the classification accuracy of MSSR-CapsNet is 97.46%,91.22%and 63.84%,respectively.In addition,the ablation experiment also verified the efficiency of the experimental method and improved the performance of the model mode.

关 键 词：胶囊网络 路由算法 多尺度表示 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]