A novel capsule network based on deep routing and residual learning

Jian Zhang,Qinghai Xu,Lili Guo,Ling Ding,Shifei Ding

DOI: https://doi.org/10.1007/s00500-023-08018-x

IF: 3.732

2023-03-22

Soft Computing

Abstract:Capsule network is a novel network structure based on vector, which has excellent performance in image classification and attracted the attention of researchers. However, the capsule network has a large amount of computation and does not perform well in the task of classifying complex datasets. Therefore, we propose a capsule network based on deep routing and residual learning. The network reduces the number of capsules to reduce the computation during routing, adds local routing layers to extract deep hidden capsule information and constructs residual connections to assist parameters training. In addition, we propose a new capsule dropout method to enhance the robustness of the network and reduce the computation of the network. We conducted several experiments on multiple benchmark datasets (CIFAR-10, SVHN, Fashion-MNIST and MNIST). The experimental results show that the proposed method achieves competitive results on multiple benchmark datasets. For example, the accuracy of 81.61% is achieved on the CIFAR-10 dataset, and the number of training parameters is reduced by 72.09%.

computer science, artificial intelligence, interdisciplinary applications