Research On Remote Sensing Image Classification Based On Capsule Network In Deep Learning

With the development of computer vision,remote sensing image classification has been widely used in many fields.The emergence of deep learning has opened up new solutions for people to study remote sensing image classification.Capsule Network(Caps Net)is a new deep learning method that uses vectors instead of scalars as the input and output of the network,and uses dynamic routing to replace the pooling operation in traditional neural networks,which can retain more posture information.Therefore,how to use the capsule network to complete the classification task has become the research focus of this article.In order to solve the problems of easy loss of feature information and complicated processing in the existing classification methods,this paper has deeply studied the capsule network,and proposed to use the capsule network to complete the classification task of remote sensing images.Firstly,the basic structure and principles of convolutional neural networks and capsule networks are analyzed.Convolutional neural networks can fully extract image features.However,when performing pooling operations,problems such as loss of spatial position information and detection target posture information will occur,resulting in low spatial resolution.The capsule network uses dynamic routing to solve this problem,so the capsule network is selected as the basic model of this article,Then on this basis,an improved remote sensing image classification model Fine-Caps is proposed on this basis.In terms of model improvement,firstly,the dense convolutional network(Dense Net)is combined with the capsule network to extract more features of the image,and then the deconvolution layer is used to replace the fully connected layer in the original model to process the complex image of remote sensing images reconstruction.In the aspect of algorithm,the dynamic routing algorithm of the capsule network is improved to an EM clustering algorithm,and the compression function is optimized.In this paper,the WHU-RS19 and NWPU-RESISC45 datasets are selected as experimental data for verification.The OA coefficient and Kappa coefficient of the Fine-Caps model on the WHU-RS19 dataset are 93.63% and 93.13%,which are 1.89% and 1.46 higher than the Caps Net model;On the NWPU-RESISC45 data set,the OA coefficient and Kappa coefficient are 93.82% and 91.75%,which are 1.81% and 1.93% higher than that of the Caps Net model.