R-CapsNet: An Improvement of Capsule Network for More Complex Data

Abstract

Convolutional neural networks (CNNs) have achieved the best performance in some fields. However, they still have some defects. CNNs need a lot of images for training; they will lose much information in the pooling layer, which reduces the spatial resolution. Facing such problems, Hinton et al. proposed a capsule network (CapsNet). Although the CapsNet has achieved the best accuracy on MNIST dataset, it has not performed well on Fashion-MNIST, Cifar-10 and other datasets. Naturally, we established an improved version of capsule network (R-CapsNet). Results have shown that when using R-CapsNet model, the loss gets decreased and the accuracy gets improved on FashionMNIST. In the meanwhile, the training parameters are reduced by nearly half. Specifically, it reduces by 4.5M. Comparisons show that our proposed model reports improved accuracy of around 0.56% over the existing state-of-the-art systems in literature. The test accuracy of R-CapsNet model is 1.32% higher than that of the original model. Furthermore, better results have been achieved on Cifar-10 with R-CapsNet model and it has easily increased by 10% compared to CapsNet.