An Improved Capsule Network Based on Newly Reconstructed Network and the Method of Sharing Parameters

摘要

The capsule network is considered as the latest technology in the field of computer vision. However, it needs a large amount of storage space due to the large amount of parameters. In this paper, we have adopted two methods to solve this problem. First, a method of sharing the parameters of capsule layer is proposed to solve the problem of too many parameters in capsule layer, which can decrease by 18% parameters compared with the previous. Second, we redesigned the structure of the reconstructed network to replace the original, reducing the network's parameters by 16%. Moreover, we combine the two methods to further reduce the parameters, which can decrease by 34%. Finally, we use the improved capsule network for MNIST handwritten digit recognition, the result is almost the same as or even slightly higher than the original capsule network, and the reconstructed images also can smooth the noise. This article provides new ideas for the future optimization methods of various capsule networks.