A novel dense capsule network based on dense capsule layers

摘要

Capsule network, which performs feature presentations for classification tasks via novel capsule forms, has attracted more and more attention. However, its performance on complex datasets has not been fully utilized. Through an in-depth exploration of Dense Convolutional Network (DenseNet), we propose a novel dense capsule network based on dense capsule layers, named DenseCaps. As far as we know, this is the first attempt to achieve a cross-capsule feature concatenations. This architecture enhances feature reuse by realizing dense connections at capsule-level, and captures different levels of detailed features to improve the performance on color datasets. Extensive experiments and ablation studies prove the proposed model achieves competitive results on multiple benchmark datasets (MNIST, Fashion-MNIST, CIFAR-10, and SVHN).