Research And Application Of Improved Classification Model Of Deep Capsule Network

Deep convolutional neural network(CNN)is widely used in image classification due to its strong feature extraction ability.The pooling layer in CNN guarantees the invariance of features by blurring the relative position relation between the target structure parts,so the classification task using CNN needs to be based on sufficient training samples.However,in the task of remote sensing scene image classification,the sample size of remote sensing scene image data set is small,which is easy to cause overfitting of the model.Moreover,the remote sensing image is characterized by the diversity of target size,the particularity of view angle and the small target.All these have brought difficulties and challenges to CNN image classification.The images in each category vary greatly in position and angle,so for the remote sensing scene data set,if the spatial attitude transformation of features can be learned,the accuracy of image classification can be improved without expanding the data set.To solve these problems,we proposed a one-shot deep capsule network.First,the deep pooling layer in CNN was removed to ensure the integrity of feature information,and capsules were introduced to extract feature spatial information,so as to learn the spatial attitude transformation matrix to reduce the the problem of target multiple scales,the particularity of perspective and the small targets of remote sensing images.Secondly,in view of the depth of the convolution network on small sample data set is not fully learning problems,the siamese network architecture in one-shot learning is proposed as an embedded verification model,and a measure learning item is added to regularize the characteristics of one-shot deep capsule network.Experimental results show that the comprehensive performance of this method is better than other methods.The model adopts the structure of twin network as embedded verification.Finally,a remote sensing image classification model prototype system is implemented,which mainly realizes the depth residual capsule network model,one-shot deep capsule network model,two remote sensing image classification evaluation indexes and three commonly used remote sensing image data sets.There are 24 figures,16 tables and 60 references in this thesis.