AlexCapsNet: an integrated architecture for image classification with noise and many categories

Abstract

Capsule network (CapsNet) is a pioneering architecture that can encode the image properties into vectors rather than scalars, addressing the limitations of traditional Convolutional Neural Networks (CNNs). This process is achieved by dynamic routing algorithms to maintain the spatial hierarchies. However CapsNet may decreases its efficacy in complex datasets, such as CIFAR-10. To address this problem, we propose AlexCapsNet architecture, in which the classic classification model AlexNet is used as the feature extraction layer. Such a layer allows CapsNet to capture deeper and more semantic features. Comprehensive evaluation with five datasets shows that AlexCapsNet has achieved improved performance compared to the baseline and other variants of CapsNet. In addition, we also introduce the Shallow-AlexNet module, revealing that deeper feature extraction layers are suitable for datasets with fewer categories and shallower feature extraction layers are more effective for datasets with many categories, owing to inherent category differences. Moreover, our investigation into the reconstruction module shows that noise in the datasets helps degrade the performance of the reconstruction module.