Weighted-capsule routing via a fuzzy gaussian model

Capsule network (CapsNet) is a novel architecture that takes into account the hierarchical pose relationships between object parts, which had achieved desirable results on image classification. EM-Routing (EM-R) used in CapsNet is the process of assigning child capsules (parts) to each parent capsule (objects) based on a level of agreement, which is similar to the fuzzy clustering process. However, CapsNet still struggles with backgrounds and the presence of noise. In this paper, a new routing algorithm based on a weighted capsule fuzzy gaussian model (WCFGM-R) and a pose loss function are proposed. The proposed algorithm aims to prohibit atypical child capsules from contaminating the parent capsules by incorporating the activations of capsules in a lower layer as weights that play the role of precision. The pose loss provides the best inter-class separation and improves the ability of pattern classification. Indeed, the experimental analyses demonstrate that CapsNet with WCFGM-R outperforms the CapsNet with EM-R in which it shows excellent results on three datasets (MNIST-bg-img, MNIST-bg-rnd, and CIFAR10).