SAR Target Image Segmentation And Classification Based On Sparse Representation And Deep Learning

As an active coherent imaging radar,Synthetic aperture radar(SAR)has the advantages of imaging at any time and any weather,which is widely used in many fields,such as exploration of mine,prevention of natural disasters and national defense,et al.In the field of military,synthetic aperture radar is mainly applied to the detection and classification of military targets,so achieving rapid and effective target classification and identification is of great significance to the national military security.In this thesis,aiming at solving the problems of poor robustness of underlying characteristics and the high complexity of the artificial design in SAR target image classification and identification,we combine the sparse representation with deep learning to get the ability of learning natural and abstract features.We propose the SAR target image classification based on deep K-means,SDAE-SVM,convolution autoencoder and probability sparse representation.The main contents are as follows:Aiming at solving the problem of high complexity of gradient descent method and many hyperparameters in CNN,we propose the method to classify the SAR target image classification with deep K-means method.Compared with convolution neural network,deep K-means method is faster and simpler,furthermore,it has only one clustering center parameter to be adjusted.By learning the kernel without supervision,deep K-means method reduces the need for the large number of labeled samples,which can also obtain abstract and natural features by stacking into a deep network.After several experiments,we can get better classification effect with this algorithm.In this thesis,we propose the SAR target classification based on SDAE-SVM method to solve the problems,including the poor robustness features of SAR image and the lower accuracy caused by using softmax classifier to fine tune the deep network.We firstly segment the image to get the target slices,which can avoid the over-fitting of the deep network,and then extract deep and robust features by using the denosing autoencoder.Finally,the linear SVM classifier is used to fine tune the network.After several experiments,we prove that our method can get good results.Facing the problem of high difficulty to initialize the dictionary with sparse representation and low accuracy of traditional sparse representation classifier.We propose SAR target image classification based on CAE and probability sparse representation.Firstly,we use CAE to extract features and reduce the dimension of the image,then regard the features as initial dictionary to represent and classify.Compared with other methods,experiments show that our algorithm can obtain higher classification accuracy.