A Survey on Capsule Networks: Evolution, Application, and Future Development

Abstract

Capsule networks have drawn increasing attention since the concept is proposed in 2011. They hold advantages over convolutional neural networks (CNNs) on modeling part-to-whole relationships between entities and learning viewpoint invariant representations. A variety of improved designs and applications related to capsule networks have been explored within the past decade. This paper presents a relatively thorough survey of literatures corresponding to the evolution of the capsule networks and their applications. Different design of capsule networks from Hinton’s group as well as other researchers are reviewed. Applications of capsule networks in various scenarios and tasks are elaborated, including hyperspectral image processing, medical image processing, facial image processing, text classification, object detection, image segmentation, few-shot learning, etc. The relation and comparison of capsule networks, CNNs, and transformers, and the trend of future development of capsule networks are discussed. We believe this review provides a good reference to researchers who are interested in capsule networks.