AlexCapsNet: an integrated architecture for image classification with background noise

Capsule networks (CapsNet) are a pioneering architecture that can encode image features
into vectors rather than scalars, addressing the limitations of traditional Convolutional Neural Networks (CNNs). This process is achieved by the Dynamic Routing algorithm and can maintain the image’s spatial hierarchies. CapsNet has demonstrated the state-of-the-art performance in simple datasets such as MNIST, but its performance degrades in more complex datasets. To solve this problem, AlexCapsNet architecture is proposed in this paper, in which the classic classification model AlexNet is used as the feature extraction layer. This allows CapsNet to capture deeper and more semantic features. The comprehensive evaluation with four datasets shows AlexCapsNet has improved performance when compared with the baseline and other CapsNet variants. Besides, our experiments on seven datasets show the reconstruction module existing in the CapsNet degrades the performance in datasets with background noise. AlexCapsNet removes the reconstruction module and therefore can adapt to these complicated datasets. Our code is available at https://github.com/BaoBao0926/AlexCapsNet.