High Resolution SAR Image Classification Based On Multi-scale Convolutional Sparse Representation

Synthetic aperture radar(SAR)is widely used in military reconnaissance,mapping,national economy fields,which can observe ground objects all day,all-weather.As the key link of SAR image application,image classification has important research significance.An important trend of SAR imaging technology is the improvement of the spatial resolution.High-resolution SAR images can provide richer ground information,but they also pose new problems for SAR image classification.Compared to low-and middle-resolution SAR images,high-resolution SAR images show the characteristics of statistical change,complex scenes.The purpose of image classification is to better understand the information of features and to classify or mark them.Because of the characteristics of high-resolution SAR image,many classification algorithms suitable for low-and medium-resolution SAR image cannot be better applied in high-resolution SAR image processing.In recent years,the convolution sparse representation replaces the product sum of the traditional dictionary and the eigenvector with the convolution sum of the two-dimensional feature map and the corresponding filter to avoid the loss of spatial information caused by the expansion of the image into one-dimensional or image segmentation,and has been widely concerned in the field of image processing.Based on the convolution sparse representation theory,this paper studies the classification of high-resolution SAR images.The main work of this paper includes the following aspects:(1)Deep learning local feature extraction model has been widely used in image processing,based on the principle of deep learning and convolution sparse representation,a high-resolution SAR image classification algorithm based multi-layer fusion convolution sparse representation is proposed.First,construct a two-layer convolution sparse coding model.The sparse feature map of the first layer is obtained by convolution sparse of the original SAR image;then,the feature map of the first layer is input into the convolution sparse model of the second layer to learn the sparse feature,and the SAR image is eventually mapped to the deep sparse value,and then the characteristics extracted from each convolution sparse layer are fused to obtain the high-dimensional feature description of the image.Finally,the SVM classifier is trained to classify the image.Through the experimental results and analysis of multiple sets of measured SAR images,the proposed algorithm can obtain better classification effect in image classification than a variety of sparse representation algorithms.(2)The high-resolution SAR image has high spatial resolution and contains rich ground feature information.Whether these ground feature information can be extracted effectively and accurately directly affects the accuracy of image classification.In order to extract the rich ground feature texture and structure information in the high-resolution SAR scene effectively,we proposes a high-resolution SAR image classification algorithm based on multi-scale anisotropic convolutional sparse representation.In this algorithm,the convolution sparse representation theory is combined with the multi-scale idea.Firstly,a set of anisotropic Gauss kernel dictionaries are learned based on the low-level feature map of SAR image,and the whole feature map is sparse decomposed to obtain the sparse representation of the spatial structure of the image in all directions.Then,based on the multi-scale Gauss convolution kernel,the feature map of each scale is superimposed to obtain the high-resolution SAR image in The context information in different scale space can effectively suppress the interference of speckle noise,obtain the multi-scale anisotropic convolution sparse feature descriptor of image,and finally train SVM classifier to realize SAR image classification.Through the simulation and analysis of three groups of measured images,it is proved that the algorithm has more accurate classification accuracy than the sparse representation algorithm in recent years.