Image classification based on quaternion-valued capsule network

In this paper, a novel quaternion-valued (QV) capsule module is designed to construct QV capsule networks for image classification. The quaternion algebra is introduced into the capsule networks to effectively capture the external dependencies and internal structural information. Moreover, the QV capsules can enhance the representation of complex information and alleviate the information loss of vanilla capsule networks. Particularly, a non-iterative quaternion routing algorithm is proposed to integrate QV capsules, considering both the membership and the consistency of QV capsules in two stages. Extensive experiments are conducted on classic image datasets, hyperspectral image datasets, and face datasets, which demonstrate that: firstly, the QV capsule network achieves higher classification accuracy, reaching 92.95% in UC Merced Land Use and 95.02% in CIFAR 10; secondly, the QV capsule module is more adaptable to different backbone networks than the vanilla capsule module; finally, the QV capsule network shows high performance with limited training samples.