A Fast Routing Capsule Network With Improved Dense Blocks

Routing algorithm in most existing Capsule Networks (CapsNets) is always a challenging problem due to its heavy computational burden. To address this issue, we propose a “fast routing” algorithm, where the high-level capsules are activated by the statistical characteristic of votes from low-level capsule vectors. In this way, the votes and their distribution are both considered, and iterations in the traditional “dynamic routing” algorithm are eliminated. The comparison experiment on MNIST dataset reveals that CapsNet with fast routing promotes time efficiency of CapsNet with dynamic routing (DR) by 71.2%, as well as improves classification accuracy by 6%. Moreover, improved dense blocks (IDB) are developed to make a powerful feature extraction, where layer position is utilized to calculate its filters to encourage feature reuse, while reducing redundant features. Finally, the proposed CapsNet with “fast routing” and IDB, namely fast routing CapsNet (FR-CapsNet), outperforms state-of-the-art capsule models in multiple benchmark datasets.