Research And Application Of Coding Capsule Network Based On Feature And Spatial Relationship

The capsule network can encode the feature information of the input image into capsules,and obtain a feature combination map of local features in the lower layer capsule to the overall features in the higher layer capsule through a dynamic routing algorithm.The dynamic routing algorithm is implemented by an iterative loop.The activation of the higher layer capsule is realized by the combination of lower layer capsules.On the one hand,this structure of the capsule network can achieve fully feature information extraction from the bottom layer to the high layer,which could realize the full use of the feature information,but it also makes it inevitably susceptible to noise features.On the other hand,the model structure and feature transfer method of the capsule network also lead to the lack of topdown spatial relationship guidance,and the utilization rate of the spatial relationship in the feature information is not high.The iterative loop algorithm for a given number of times also makes the feature optimization process of combining lower layer capsules to activate higher layer capsules inefficient,which leads to slow convergence and poor accuracy of the model when the capsule network processes input images with complex features and spatial information.Therefore,this paper proposes a new capsule network based on feature and spatial relationship coding,which introduces the top-down spatial relationship information for the capsule network to optimize the capsule network structure.On the one hand,the feature and spatial relationship extractor are set to abstract the feature and spatial relationship respectively,and the interference of noise information is weakened.On the other hand,the feature and spatial relationship encoder are set up to abandon the iterative optimization method,and the process of finding the optimal feature combination is added.Backpropagation accelerates the convergence of the capsule network and optimize the feature and spatial relationships.In addition to this,this article also explores the possibility of using a deconvolution layer instead of a fully connected layer to construct a reconstructed unit.Finally,compared with the standard capsule network and its various variant models,the feature and spatial relationship-encoding capsule network proposed in this paper has achieved significantly better performance on the Fashion-MNIST and CIFAR-10 data sets.In addition,compared to some mainstream convolutional neural networks,the feature and spatial relationship-encoding capsule network have also achieved fairly competitive accuracy on Fashion-MNIST with a simpler network structure,thanks to its efficient spatial relationships processing power.In view of the above advantages of the feature and spatial relationship encoding capsule network,at the same time,the lung node recognition model based on the convolutional neural network has the problems of complex model structure,slow convergence,and high false positive rate of the recognition result.In this paper,the feature and spatial relationship encoding capsule network is used to replace the traditional Convolutional neural network,combined with the image segmentation network U-net,constructs a hybrid model for lung node recognition tasks in lung CT images.At the same time,in view of the problem that the false positive rate of suspected nodes in the recognition process is too high,this paper has carried out various performance optimizations for U-net,and applied various technical methods such as data enhancement,threshold setting,and model fusion.Finally,after introducing various evaluation indicators,the lung node recognition model based on feature and spatial relationship encoding capsule network has achieved better accuracy,faster convergence speed and lower false positive rate of suspected nodes on the LUNA-16 dataset than the lung node recognition model based on residual neural network performance.