Research On SAR Image Classification Algorithm Based On Capsule Network

Synthetic Aperture Radar(SAR)is a microwave remote sensing imaging system,which is active and can collect high-resolution,all-day images.It is unique in disaster monitoring,environmental monitoring,geological survey and military.It can play a role that other remote sensing methods can’t play.In recent years,with the continuous development of deep learning,the automatic target recognition(ATR)method of SAR images has gradually developed from the traditional statistical-based method to the deep learning-based method.Among them,the ATR based on the convolutional neural network research has made great progress.However,there are still many challenges in SAR image recognition,such as complex SAR imaging mechanism,difficulty interpreting imaging results,speckle noise in the images,and scarcity of labeled data sets for training.The main research works and innovation results of this paper are as follows:(1)A SAR image classification algorithm based on the attention mechanism of the capsule network is proposed.The attention capsule network model is composed of channel attention module,spatial attention module,convolutional layer,primary capsule layer,SAR capsule layer,and reconstruction network module.First,the preprocessed SAR image passes through the convolutional layer to extract the primary features.Then the primary features are sent to the channel attention module and the spatial attention module to extract more accurate and discriminative high-level features,and these high-level features are sent to the primary capsule layer.Then these features are input into the SAR capsule layer by the dynamic routing,so as to obtain the position and spatial relationship between different entities in the SAR image.Finally,the output of the SAR capsule layer is input to the reconstruction network module to reconstruct the SAR image,and participates in the loss calculation with the original SAR image.Experiments show that this method achieves better results on the classic SAR image classification data set MSTAR,which verifies the effectiveness of this method.(2)A SAR image classification algorithm based on dense capsule network with limited data is proposed.Aiming at the problem of the shortage of samples in the SAR image annotation data set,a dense capsule network model is proposed.The model is composed of three dense primary capsule modules,a SAR capsule layer,and a reconstruction network module.The model uses three dense capsule layers to extract the deep position and spatial features from different levels in the SAR image,and then fuse the extracted features at different levels.The fused features are input into the SAR capsule layer for classification by the dynamic routing.Finally,the output of the SAR capsule layer is input to the reconstruction network to reconstruct the SAR image,and participates in the loss calculation with the original SAR image.The first method is mainly aimed at extracting discriminative and robust SAR image features to achieve the effect of improving the classification accuracy.The second method focuses on the classification of SAR images on small data set.The network needs to extract deeper and more levels of image features.Therefore,the network structure model of the second proposed method is deeper and wider.In order to improve the accuracy of SAR image classification with limited training samples.