Yuan Hailei, Ren Qiang
School of Yonyou Digital and Intelligence, Nantong Institute of Technology, Nantong, China.
School of Computer Science and Technology, Tongji University, Shanghai, China.
PLoS One. 2025 Jul 30;20(7):e0329202. doi: 10.1371/journal.pone.0329202. eCollection 2025.
Capsule Networks (CapsNets) have demonstrated an enhanced ability to capture spatial relationships and preserve hierarchical feature representations compared to Convolutional Neural Networks (CNNs). However, the dynamic routing mechanism in CapsNets introduces substantial computational costs and limits scalability. In this paper, we propose a divide-and-conquer routing algorithm that groups primary capsules, enabling the model to leverage independent feature subspaces for more precise and efficient feature learning. By partitioning the primary capsules, the initialization of coupling coefficients is aligned with the hierarchical structure of the capsules, addressing the limitations of existing initialization strategies that either disrupt feature aggregation or lead to excessively small activation values. Additionally, the grouped routing mechanism simplifies the iterative process, reducing computational overhead and improving scalability. Extensive experiments on benchmark image classification datasets demonstrate that our approach consistently outperforms the original dynamic routing algorithm as well as other state-of-the-art routing strategies, resulting in improved feature learning and classification accuracy. Our code is available at: https://github.com/rqfzpy/DC-CapsNet.
与卷积神经网络（CNN）相比，胶囊网络（CapsNets）已展现出更强的捕捉空间关系和保留层次特征表示的能力。然而，CapsNets中的动态路由机制带来了巨大的计算成本并限制了可扩展性。在本文中，我们提出了一种分治路由算法，该算法对主胶囊进行分组，使模型能够利用独立的特征子空间进行更精确高效的特征学习。通过对主胶囊进行划分，耦合系数的初始化与胶囊的层次结构对齐，解决了现有初始化策略中要么破坏特征聚合要么导致激活值过小的局限性。此外，分组路由机制简化了迭代过程，减少了计算开销并提高了可扩展性。在基准图像分类数据集上进行的大量实验表明，我们的方法始终优于原始动态路由算法以及其他当前最先进的路由策略，从而提高了特征学习和分类准确率。我们的代码可在以下网址获取：https://github.com/rqfzpy/DC-CapsNet 。
