High Resolution SAR Image Classification Based On Deep Learning

Synthetic Aperture Radar(SAR)can continue to work in all weathers,and because the scattering information and texture information of the SAR image are very abundant,it is not only applied in the military field,such as object classification and target detection,target recognition and tracking,etc.,but also widely used in commercial applications,such as city planning,mineral detection,agricultural production.It is of great application value.With the continuous development of SAR technology,the resolution of image is higher and higher,and the requirement of automatic interpretation of SAR images is also increasing.How to interpret SAR image data accurately and quickly is a hot issue to be solved.In recent years,the development of deep learning technology has been very rapid,which has attracted great attention from industry and academia.The technology of deep learning has achieved great success.In this paper,the main work is to apply the fully convolutional networks method in deep learning to the field of high resolution SAR image classification,and propose some improvements,as follows:1.We studied the fully convolutional networks,which is a variation of the convolutional neural networks,and emphatically discusses the basic composition and its algorithm,then applies the fully convolutional networks in the field of high resolution SAR image classification.We proposes a tiny fully convolutional networks structure,called Tiny FCN,and use it for SAR image classification.It replaces traditional “patch” classification method in convolutional neural networks and SVM to “pixel” classification,which maintains high accuracy and speeds up the classification process.Experimental results show that our method is superior to the Res Net-18 with FCN.2.Single branch fully convolutional network lacks of multi-scale information.This paper presents a high resolution SAR image classification method based on multi-scale fully convolutional networks.We design a multi-scale convolutional structure,bring it into the tiny fully convolutional networks,and replace the 3rd convolution layer.The multi-scale feature information of high resolution SAR image is extracted through multiple channels,and then the feature fusion is made.The experimental results show that the multi-scale structure of this paper further enhances the classification accuracy and it is still very fast,and it's better than the pyramid pooling.3.Considering that the fully convolutional network only utilizes the gray information of the original SAR image,the feature is deficient.In this paper,a multi-scale fully convolutional network based on feature combination is proposed.The main idea is to enrich the input of the multi-scale fully convolutional network.Firstly,the multi-direction Gabor texture feature of the original SAR image is extracted.And then,the information is combined with the original image in the input layer.Finally,it is fed into the multi-scale fully convolutional network.The experimental results show that the method improves the classification accuracy further.