The Analysis Between Traditional Convolution Neural Network and CapsuleNet

Convolution neural networks(CNNs) have made a series of breakthroughs in the fields of autonomous driving, robotics, medical treatment and so on. The powerful feature learning and classification abilities of CNNs have attracted wide attention of scholars. With the improvement of the networks structure and the increasing of the networks depth, the performance of tasks such as image classification, detection, recognition, and segmentation has been improved greatly. Recently, CapsuleNet, which is different from CNNs, has been proposed. In CapsuleNet, vectors replace traditional scalar points to better characterize the relationships between feature information and acquire more attributes of things. In this paper, the basic structure and algorithm principle of classical CNN models were described firstly, and then the basic neural network, network structure, parameter update and distribution process for the CapsuleNet model were analyzed and summarized. Finally, several group of training experiments were conducted to compare and analyze the image classification and feature expression canabilities of CansuleNet and CNN models.