A Robustness Enhanced Capsule Network with Regularization

This paper aims to guarantee the robustness of capsule networks (CapsNets) by regularization in the construction of primary capsules and higher-level capsules. A novel CapsN et, namely robustness enhanced CapsNet with regularization (RE-CapsN et) has been put forward by exploiting the equivariant spatial relationship and the invariant descriptors of parts. The main idea of RE-CapsNet is three-fold: 1) the regularization for poses and descriptors of capsules combined with attention modules is introduced to establish the primary capsules; 2) the utilization of regularization contributes greatly to the robustness, and the adoption of attention modules and regularization gives rise to good interpretability of RE-CapsNet; 3) the comparison experiments on several datasets reveal the RE-CapsNet generalizes better on unfamiliar viewpoints with affine transformations.