Improvement Of Capsule Network And Its Application In Image Generation

The capsule network was proposed in October 2017 which has caused a huge sensation in the academic world as the latest technology in the field of computer vision.The capsule network expresses the relation of part and whole in the form of vectors,which not only can express the image with the intensity of the feature response,but also can characterize the direction,location and other information of the image feature.The update between the two layers vectors of the capsule network needs to use the dynamic routing algorithm to replace the max-pooling method used in traditional convolutional neural networks,so as to avoid the loss of accurate position information of images.However,because it hasn't been long since the capsule network was proposed,which still has many imperfections.In this thesis,we focuses on the studies of the capsule network,including its network models,training methods and applications,and meanwhile in order to further improve the practicality of the capsule network,the optimization methods of convolutional neural networks are combined effectively.The specific research contents are as follows:(1)Capsule networks already have high classification accuracy on simple data sets such as handwritten digits.However,in order to enable them to be practically applied into mobile devices such as Android,the less network parameters the better.In order to reduce the complexity of the network,in this thesis we propose a training method of sharing parameters between capsule layers and redesign the reconstructed network,and then combine deconvolution network with the optimization method of Batch Normalization to further reduce the amount of parameters in the network.Finally,experiments conducted on the MNIST handwritten digital dataset can greatly reduce the parameters of the network in the meantime can maintain the original recognition accuracy and balance the relationship between recognition accuracy and network complexity.(2)In view of the fact that the recognition accuracy of the capsule network on color images is not very high,so a more effective improved version of the capsule network-EfCaps proposed,which is mainly composed of three sub-networks,namely feature extraction network,dynamic routing network and reconstructed network.The improvement of the feature extraction network is to use the convolution kernel of different sizes in parallel with convolution to fully extract features and obtain feature combinations of different receptive fields;the improvement of the dynamic routing network is to use the 1×1 convolution kernel to do downsampling of the capsules,which can remove redundant capsules and reduce the complexity of the network;the reconstructed network uses the parallel deconvolution is the same kernel size as the feature extraction network to recover the image,and there are two skip connections between the feature extraction networks which can reconstruct a more realistic image and can facilitate the recognition of the output capsule vector.Finally,through the experiments on the CIFAR-10 dataset,the improved capsule network proposed in this thesis is 5% higher in recognition accuracy than the original network,and its speed is far superior to the original network's.(3)The output of the capsule network is in the form of a vector,which is capable of mapping specific categories of data onto a potential vector space,and through which the image data at the pixel level can be well reconstructed.Aiming at this characteristic,this thesis applies the capsule network to data generation and proposes a conditional auto-encoding generative adversarial networks based on capsule network,which effectively combines the advantages of auto-encoder and generative adversarial networks in data generation.And the capsule network is used as a conditional auto-encoder to map sample data to potential feature representation spaces.Then there are two pairs of generative adversarial networks,one pair is for the potential vector space,the generator samples the evenly distributed noise to map it to the potential vector space of the capsule network representation,and the discriminator discriminant vector is derived from the real vector space of capsule network representation.The other pair is for the sample space,the generator maps the generated vector of the potential vector space to the sample space,and the discriminator discriminates whether the sample is from the real sample space.Through the joint training of these networks,experiments on the MNIST and FashionMNIST datasets prove that the conditional generation model proposed in this paper can generate more stable and clear samples without problems such as mode collapse,and the use of a capsule network as a conditional auto-encoder can effectively map different classes of samples to potential vector spaces that can be distinguished.In this thesis,the model and training algorithm of the capsule network are investigated and improved,and it can applied to the image generation,which opens up a new application field for the capsule network.At the same time,it also provides a reference direction for the future research and development of the capsule network.