Research And Application Of Dynamic Routing Based On Capsule Networ

Capsule Network encodes the entity and spatial relationship of images into vector capsules.The effective information is extracted from low-level capsules into high-level capsules by dynamic routing.The iterative update method of dynamic routing can make full use of the feature information of images,but it also has limitations.On the one hand,the fully connected structure cannot effectively distinguish the contribution of features,and the number of iterations specified artificially makes the model more vulnerable to noise in complex scenes;on the other hand,dynamic routing independent of backpropagation only relies on coupling coefficients to filter feature,and its iterative update will also suppress the support of multiple advanced capsules by the same key feature,resulting in the loss of effective features.In addition,the feature extraction of a single convolutional layer is not only insufficient,but also makes the capsule contain more redundant feature,which will affect the convergence speed and classification accuracy of the model.To address the above problems,this thesis proposes two models to explore the feasibility of Capsule Network in theory and application.When dealing with complex images,the Capsule Network has poor performance caused by the dynamic routing,which cannot filter the noise information effectively.Aiming at the problem,this thesis proposes a capsule graph network model based on dot product attention(DPA-Caps Graph).To effectively extract key features in images,channel attention is added to the feature extraction module,and large weights are given to effective features while suppressing the negative effects of noise.In order to obtain the dependencies between capsules in the same layer,the point product attention was used to build a capsule graph structure between the obtained high-quality primary capsules.The Laplacian feature mapping was used to obtain high-level capsules,so as to improve the model’s performance in the same layer of capsules.DPA-Caps Graph was trained by back-propagation,which makes up for the defect that dynamic routing needs to be independently updated between capsule layers.To maintain the length advantage of the higher capsule,the hyperparameters in the squash activation function are appropriately adjusted.The experimental results show that DPA-Caps Graph not only reduces the number of parameters,but also has better performance on four benchmark datasets such as CIFAR10.Due to the small amount of sample data and complex structure in real cerebral infarction MRI images,the classification accuracy of the Capsule Network needs to be improved.Aiming at this problem,a composite routing Capsule Network model based on the residual structure(DPA-Caps Routing)is proposed based on the DPA-Caps Graph.Through the residual structure,the effective information of the cerebral infarction image in the spatial and channel dimensions is fully extracted.The identity mapping added between the residual blocks not only extracts complex and abstract high-level features but also fuses important low-level features into vector capsules.The residual structure optimizes the capsule quality and realizes the sparsity expression of the model.To further improve the performance of the model,the feature matching degree between different layers of capsules is added after the same-layer capsule graph routing.The effective features from multiple angles were fully captured by the multi-level routing process.DPA-Caps Routing not only outperformed other Capsule Network variants on the benchmark dataset but also achieved a classification accuracy of 98.75% in the classification of cerebral infarction MRI images,which further verified the effectiveness and generalization of the proposed model.