Capsule Network with Inception Structure for Image Classification

Convolutional Neural Networks (CNNs) are popular in deep learning, especially in computer vision. But CNNs are poor at extracting spatial information, such as position and direction of entities in an image. A recently proposed model called Capsule network could effectively learn the spatial relations among entities, but it has weak feature extraction ability. The inception structure could extract multi-scale features. So this paper modifies the inception-v1 structure and adds it into Capsule network to strengthen the feature extraction ability. In the modified inception-v1, the max pooling branch is removed to reduce the loss of feature information. And Batch Normalization (BN) layer is added after each convolution to accelerate convergence and reduce over fitting. To organize categories information, a fully connected layer is added after digit Capsules. This paper conducted experiment on two public datasets, and the results show that the proposed model outperforms the original model in accuracy.