Research On Capsule Network Deep Learning Method And Its Application In Hyperspectral Image Classification

Hyperspectral image classification has always been a research hotspot in the field of remote sensing.With the improvement of hyperspectral imaging technology,the resolution of hyperspectral image data is also getting higher.They bring richer spectral information and spatial structure information,and at the same time put forward higher requirements on the classification model.At present,the deep learning model based on the convolutional neural network is one of the most commonly used models for hyperspectral image classification.However,the traditional convolutional neural network model has a weak ability to fuse spatial and spectral information.In addition,it has a serious averaging effect and is not sensitive enough to the difference of spatial and spectral information.The capsule network uses the ability of the capsule vector to express features and the ability of dynamic routing algorithms to integrate features to improve the classification performance of the model.Firstly,this article deeply researches the theory and classic model of convolutional neural network,by comparing the classic convolution structure and analyzing both the advantages and disadvantages of convolution.On the basis of the capsule network,the convolution structures are improved to extract multi-scale features and shallow features,which are able to alleviate excessive smoothing.At the same time,based on the capsule vector expression and sub-capsule division between channels,the network expresses the diversity of the data space structure for subsequent construction.As the control test results of subsequent models,a variety of classic models were tested on five hyperspectral image data sets of Indian Pines,Salinas,Tea Farm,Pavia University and Xiong’an.Secondly,analyzing the characteristics of high-resolution hyperspectral images,we design a convolution module and construct a hyperspectral image classification model based on spectralcapsule and spatial-capsule network,respectively.By vector-to-vector or matrix-to-matrix information transfer it also improves the interference of the diversity of ground features with the high resolution on classification,while express the local overall relationship of the data.The models are verified on five hyperspectral image data sets,which prove the effectiveness of the proposed capsule network in hyperspectral image classification,and the classification accuracy is improved on five hyperspectral image data sets.Finally,the attention mechanism is embedded in the capsule network to improve the model’s ability to express features.The attention mechanism is embedded in the convolution structure to reduce the fragmentation characteristics and improve the convergence speed of the model.The capsule spatial attention is used to replace the affine transformation matrix to maintain the spatial structure characteristics of the capsule matrix.The capsule network embedded with the attention mechanism converges fast during training,the model accuracy is higher,and the generalization is stronger.Through comparative tests on five hyperspectral image data sets,the capsule network with attention embedded in the convolutional structure has high practicability.The capsule structure that uses the attention mechanism to achieve affine transformation has good classification results,which provides new ideas for capsule network research.