An Improved Capsule Network Based on Capsule Filter Routing

Abstract

Capsule network (CapsNet) is a novel type of network that can retain spatial information, because each capsule can integrate more information than scalar-output features. However, the CapsNet learns all the features in the input image due to the lack of pooling operation, and there is no connection between different layers in the multi-layer network structure. In this paper, we propose an improved capsule network (CapsNet) based on Capsule Filter Routing (CFR) to address this problem, namely CFR-CapsNet. Firstly, we propose a new routing method called CFR for filtering capsules based on capsule activation values, which can speed up the operation of the model, and then introduce a self-attention mechanism to improve the performance of the primary capsule in the capsule space. Furthermore, in the multi-layer network structure, we transmit the information of the classified capsule with the largest activation value in the previous capsule layer to the primary capsules of the next layer, which improves the relevance of the overall structure. We conduct experiments on Fashion-MNIST, SVHN, and CIFAR-10/100, and the experimental results show that our method can improve the performance of the CapsNet more effectively than other state-of-the-arts.