TITLE:
Lightweight Capsule Network Based on Weight Sharing and Top-K Routing
AUTHORS:
Dazhong Mu, Ran Li
KEYWORDS:
Capsule Network, Lightweight Model, Deep Learning, Weight Sharing, Dynamic Routing, Attention Mechanism
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.4,
April
2,
2026
ABSTRACT: Capsule networks can effectively model the spatial hierarchical relationships among features; however, the dynamic routing mechanism introduces a large number of parameters and high computational complexity, which limits their practical deployment in real-world engineering applications. To address these issues, this paper proposes a lightweight capsule network named WTCaps, based on weight sharing and Top-K routing pruning. Specifically, a weight-sharing strategy is introduced in the capsule transformation stage to significantly reduce the number of model parameters. Meanwhile, a Top-K routing pruning mechanism is employed to perform dynamic routing only among a small subset of highly relevant higher-level capsules, thereby effectively decreasing computational complexity. In addition, a lightweight spatial attention mechanism is incorporated to enhance the representation of critical regions, improving classification performance while maintaining model efficiency. Experimental results on multiple public datasets demonstrate that WTCaps achieves classification accuracy comparable to or better than existing capsule networks, with substantial reductions in both parameter count and computational cost. Furthermore, the proposed model exhibits strong robustness in rotated digit recognition and overlapping digit recognition tasks. This study provides a feasible solution for the efficient design and practical application of capsule networks.