Research On Image Recognition Algorithm Based On Capsule Network

In 2018,capsule network was listed as one of the top ten growth technologies of artificial intelligence by Chinese Institute of Electronics.Capsule network is a new deep learning method,which was proposed by Geoffrey Hinton,the father of deep learning.This new structure of neural network has attracted extensive attention in the field of machine learning,and may bring revolutionary significance to artificial intelligence technology in the future.The capsule in the capsule network consists of a group of neurons,that is to say,a group of neurons are vectorized.The length of the vector is used to represent the probability of the existence of an entity or a part of an entity,and the direction of the vector is used to represent their various attributes,such as position,direction,size,deformation,speed,color,etc.Vector can better represent the image details and the spatial relationship of entities,which greatly makes up for the shortcomings of convolutional neural network.However,due to the effectiveness of this new network in different tasks to be verified,the development of capsule network is still in its infancy.Therefore,this dissertation focuses on the capsule network,studies its network structure,recognition speed and other applications.In addition,some optimization methods of convolution neural network are used for reference,and some shortcomings of capsule network are improved synthetically,which makes it more convenient to use in related fields.The specific research contents are as follows:1)One of the innovations of capsule network is to propose a vectorized capsule,which can express the pose information of object better than the scalar of traditional neural network,and can learn more robust representation.Through visualization experiments,this dissertation study how the vector inside the capsule affects the posture of the entity,which shows that it is a potential direction.2)Although the capsule network has achieved the best accuracy on MNIST data set recognition,it has not performed well on the fashion clothing image data set.In this dissertation,we improve the capsule network,discuss the reason why it can't recognize the more complex data,and through the improvement of feature extraction and training methods.Specifically,it mainly uses the idea of inception module for reference and introduces batch normalization in the reconstructed layer.We get better recognition accuracy than the original model on Fashion-MNIST test data set.In addition,in order to prove the effectiveness of our experiment,we also verify it on CIFAR10 data set.The results show that it is better than MNIST baseline network when the parameters are less than the original network,and the accuracy of the model is about 10% higher than that of the original capsule model.3)Because the recognition performance of capsule network on color image is not ideal,and a large number of parameters make it difficult to apply in practice.Based on these observations,a more accurate capsule network is designed,which we call PdCaps.The network consists of three subnetworks.The first is feature extraction network,we design two convolution layers with different convolution kernel sizes.The second is parallel convolution and dynamic routing network,we propose a more effective dynamic routing mechanism than the original one.In addition,parallel convolution is performed through two different convolution kernel to obtain PrimaryCaps.The third is the decoding network,which can avoid over fitting of capsule network by adding decoding layer.We add deconvolution layer to reduce the parameters of this part without affecting the overall performance of the model.Then,PdCaps is used to identify CIFAR10 data set,which not only improves the accuracy by more than 12%,but also reduces the number of parameters in the model.This dissertation provides a new direction for the design of capsule network which is closer to practical application.In this dissertation,the network model,training algorithm and other aspects of the capsule network are studied and improved.And its application in image recognition is studied,which lays a foundation for the application of the capsule network in the actual application scenarios.Moreover,it provides a reference direction for the future research and development of the capsule network.