Research On Capsule Network Models For Image Classification Task

Due to the operation of pooling and convolution,convolution neural network possesses translation invariance,which has made it the preferred algorithm in the field of image classification.However,convolution neural network ignores translation covariance to some extent.Moreover,the classification process of convolution neural network is similar to a black box operation,but the explanation is poor.In addition,it ignores the spatial relationship of the internal attributes of the image,which is not in line with the visual logic of the human eyes.In contrast,as a new type of neural network,capsule network uses vectors to represent features,so that it possesses translation covariance.And it uses parameter matrix to encode the spatial relationship within the target object,so that it can be explained.However,the capsule network has some defects: 1.Capsule network does not perform well in the classification of complex datasets,and the amount of computational and parameters in the routing process is too large;2.Capsule networks cannot increase classification performance by stacking routing layers.In view of these problems,the following research work has been carried out in this thesis:1.In view of the poor performance of capsule network in the classification of complex datasets and the huge computational overhead in the routing process,a capsule network model based on multipath feature(MCNet)is proposed.MCNet includes a new capsule feature extractor and a new capsule pooling method.The capsule feature extractor extracts the features of different levels and different locations from different paths in parallel,and then encodes the features as capsule features containing more semantic information,so as to improve the representation ability of the model.The capsule pooling method selects the most active capsules at each location of the capsule feature map,and uses a small number of capsules to express the effective capsule features,which greatly reduces the number of parameters and the amount of calculation in the routing process.Experimental results show that this algorithm improves the classification performance of capsule network on complex datasets and reduces the calculation cost of routing process.2.In order to solve the problem that capsule network can not improve the classification performance by stacking routing layer,a capsule network model based on deep routing and residual learning(RCNet)is proposed.First,RCNet introduces local routes with minimal parameters to solve the computational barrier of stacked routing layer.Then,a deep routing structure was constructed by stacking local routing layers and dynamic routing layers to explore the features of deep capsules.Finally,residual learning is introduced into the deep routing structure to help construct more effective gradient flow and assist parameter training.In addition,RCNet includes a new capsule dropout method that modifies the dropout method applied to scalar features,which reduces computational by randomly masking the overall features of some capsule types and improves classification performance by enhancing model generalization capabilities.The experimental results show that the combination of deep routing and residual learning can improve the classification performance of capsule networks by stacking routing layers.There are 36 figures,13 tables and 87 references in this thesis.