Capsule Networks in Computer Vision Applications

Capsule networks represent a paradigm shift in deep learning by grouping neurons into capsules that encode both the probability of a feature’s presence and its instantiation parameters, such as pose or orientation. In contrast to conventional convolutional neural networks, which often discard spatial hierarchies through pooling, capsule networks preserve part–whole relationships and exhibit equivariance to affine transformations. Central to their operation is a routing mechanism that assigns lower-level capsules to higher-level ones based on agreement, enabling more robust generalisation to novel viewpoints with fewer parameters. Since their introduction, capsule networks have been applied to image classification, segmentation, object detection, medical imaging and remote sensing, demonstrating enhanced resilience to occlusion, deformation and background variation. Recent efforts have focused on optimising routing efficiency, integrating attention modules and extending capsule representations into complex-valued and probabilistic domains, thereby fostering more interpretable and data-efficient vision systems.

Research from Nature Portfolio

Recent studies have advanced the efficiency and accuracy of capsule networks through novel attention-based routing strategies. One investigation replaced iterative dynamic routing with a non-iterative, self-attention routing algorithm, achieving state-of-the-art performance on multiple benchmarks while reducing parameter count by over 98 %. Another work introduced a dual attention mechanism, applying attention at both convolutional and capsule layers, which markedly improved convergence speed and classification accuracy on datasets including MNIST, CIFAR-10, smallNORB and COIL-20, and delivered superior image reconstructions by emphasising salient spatial features during network training.

Capsule Networks in Computer Vision Applications publication trend

The graph below shows the total number of articles in capsule networks in computer vision applications across all publications each year (not limited to Nature Index journals).

Technical terms

Capsule: A group of neurons arranged as a vector or matrix that encodes both feature presence and instantiation parameters such as pose or deformation.

Dynamic routing: An iterative agreement‐based algorithm that directs outputs from lower‐level capsules to appropriate higher‐level capsules based on consensus of predictions.

Attention mechanism: A strategy that assigns adaptive weights to features or capsules, emphasising informative elements and suppressing irrelevant ones.

Equivariance: A property in which transformations applied to the input yield corresponding transformations in the internal representation rather than invariant responses.

References

1. Efficient-CapsNet: capsule network with self-attention routing. Scientific Reports (2021).
2. DA-CapsNet: dual attention mechanism capsule network. Scientific Reports (2020).
3. Capsule Network with Its Limitation, Modification, and Applications—A Survey. Machine Learning and Knowledge Extraction (2023).
4. No routing needed between capsules. Neurocomputing (2021).
5. Capsule Routing via Variational Bayes. Proceedings of the AAAI Conference on Artificial Intelligence (2020).
6. Cv-CapsNet: Complex-Valued Capsule Network. IEEE Access (2019).