Multi-stage Attention-based Capsule Networks for Image Classification
Aiming to address the inadequate feature extraction problems in the traditional capsule networks(CapsNets),a multi-stage attention-based CapsNet model is proposed in this paper for image classification.Firstly,spatial attention(SA)and channel attention(CA)are used to extract effective features in the convolutional layer from low-level features and high-level features,respectively.Then,attention mechanism based on vector direction is introduced into the dynamic routing layer to enhance the focus on the important capsules,thereby improving the prediction accuracy of the low-layer capsules to the high-layer capsules.Finally,the comparison experiments on im-age classification are carried out on five public datasets.The experimental results show that the proposed CapsNet outperforms other CapsNets at the classification accuracy and the robustness,and its shows a good performance on the image reconstruction for affine images.
万方数据
维普
