基于迁移学习的注意力胶囊网络

胶囊网络(Capsule Network,CapsNet)通过运用胶囊取代传统神经元,能有效解决卷积神经网络(Conventional Neural Network,CNN)中位置信息缺失的问题,近年来在图像分类中受到了极大的关注.由于胶囊网络的研究尚处于起步阶段,因此目前大多数胶囊网络研究成果在复杂数据集上表现的分类...

Attention Capsule Network based on Transfer Learning

In recent years,Capsule Network(CapsNet)has received great attention in image classification because it replaces traditional neurons with capsules and overcomes the defects of losing position information in Convolutional Neural Network(CNN).Since the research of CapsNet is still in its infancy,most research results of CapsNet have poor classification performance on complex datasets.To solve this problem,a new capsule network is proposed to complete the image classification task,named Attention Capsule Network based on Transfer Learning,by improving the feature extraction network through transfer learning and integrating the attention module.Firstly,a 9-layer feature extraction network with the ELU activation function is used to extract features;secondly,the parameters obtained from the feature extraction network training on the ImageNet dataset are used on the CIFAR10 dataset through Transfer Learning;thirdly,the attention module is stacked after the feature extraction network to extract key features.Finally,experiments on public datasets including CIFAR10,SVHN,MNIST,and FashionMNIST show that the proposed Attention Capsule Network based on Transfer Learning can achieve ideal classification accuracy on both simple and complex datasets.