关键词: 胶囊网络, 深度学习, 动态路由, 胶囊池化, 反卷积重构

Abstract: Abstract: In view of the problems that capsule network (CapsNet) does not perform well in the classification of complex datasets and thers is a large amount of computation in the routing process, the capsule network based on multipath feature (MCNet) was proposed. A novel capsule feature extractor and a novel capsule pooling method were included in MCNet. Tthe features of different layers and locations were extracted in parallel from different paths by the capsule feature extractor, and then the features were encoded into capsule features that contain more semantic information. In the capsule pooling method, the most active capsules at each position of the capsule feature map were selected, and the effective capsule features were represented by the small number of capsules, Experiments were carried out on multiple datasets (CIFAR10, SVHN, Fashion-MNIST, MNIST). On the CIFAR10 dataset, the classification accuracy of MCNet is 79.27%, and the number of trainable parameters is 5.88M. Compared with CapsNet, the classification accuracy of MCNet is improved by 8.7%, and the number of parameters is reduced by 39.8%. The experimental results show that, MCNet not only improves the classification accuracy but also reduces the number of trainable parameters.

Key words: Keywords: capsule network, deep learning, dynamic routing, capsule pooling, deconvolution reconstruction