Modified Capsule Network for Object Classification

摘要

The recognition of images in complex scenes is essential to intelligent unmanned systems. The CapsNet performs well on MNIST datasets with overlapping numbers, but it has too many parameters on real scene datasets. In this paper, we proposes three methods to reduce its excessive parameters: (1) proposing the CapsNetPr network, in which the shallow feature extraction network is introduced, to reduce the data dimension of the input capsule layer. (2) utilizing the method of decomposing the transformation matrix to reduce space consumption and time consumption. (3) sharing the transformation matrix on the same location to reduce the number of matrices in the low-level capsule layers. The study successfully reduces the number of parameters of the capsule network and accelerates training and testing at the same time, which is of great value to the promotion and use of the capsule network.