Optimization Of Dynamic Routing Algorithm In Capsule Network

Since the capsule network was proposed,it has quickly become a research focus of scholars at home and abroad due to its excellent feature expression ability and strong robustness,and has been applied to many fields such as text classification,image classification,and behavior recognition,etc.The number of capsules,the structure of the network,and the combination with other algorithms have become the main research directions in the field of capsule networks.However,the high computational cost of dynamic routing in the capsule network currently limits its application prospects,and there are still few methods to optimize dynamic routing.The capsule network has two network structures: vector capsule network and matrix capsule network,both of which have different capsule types and dynamic routing methods.In the vector capsule network,a one-element vector capsule is used,and dynamic routing uses a nonlinear function;in a matrix capsule network,a two-element matrix capsule is used and an activation probability is added,and the dynamic routing uses the EM algorithm.Aiming at the problems of dynamic routing in the two network structures,this paper optimizes and improves the two dynamic routing rules.In the vector capsule network,in order to solve the problem that the original clustering algorithm in dynamic routing is sensitive to the selection of initial clustering center,the density peak clustering(DPC)algorithm is used to optimize the original clustering algorithm.And we propose DPC-Caps Net model to improve the overall performance of dynamic routing algorithms.The experimental results of the DPC-Caps Net model based on the Tensor Flow framework show that the capsule network structure combined with the DPC algorithm has fast convergence speed and high classification accuracy on both MNIST and Fashion-MNIST data sets.In the matrix capsule network,aiming at the problem of excessive parameters in dynamic routing,optimize from the perspective of non-parametric clustering,combine KDE algorithm with dynamic routing,propose a matrix capsule network algorithm based on KDE optimized dynamic routing,and apply it to the identification and classification of melanoma in dermoscopy images.Experiments based on the Tensor Flow framework combined with multiple evaluation indicators show that the average accuracy of the recognition algorithm on the ISIC-2017 data set reaches 98.2%,which is a great improvement compared to other network structures.In summary,this paper studies the dynamic routing algorithms in the vector capsule network and matrix capsule network.The overall performance of the model is further improved by adding the DPC algorithm and the KDE algorithm respectively.The experimental results verify the effectiveness of the proposed algorithm.It also proves the application prospects of the capsule network.