SAR Image Target Recognition Methods Based On Joint Sparse Representation And Deep Learning

Synthetic aperture radar(SAR)imaging technology has the advantages of all-day,all-weather,high resolution,long detection distance,penetration of clouds and vegetation.It is an advanced method of acquiring radar information and has important military and civilian values.However,the characteristics of SAR images will change significantly with different configuration conditions,making the target recognition task of SAR images not as easy as optical images.In addition,with the development of SAR imaging technology,the amount of data contained in radar images is getting larger and larger,which makes it difficult to interpret SAR images.How to identify objects in radar images and classify targets is an urgent problem in radar image interpretation.This thesis mainly studies the methods of recognizing targets in SAR images by applying joint sparse representation model and deep learning technology.The joint sparse representation model can not only represent a single task,but also integrate the connections of multiple tasks,giving play to their complementary advantages.Deep learning,as a machine learning method proposed in recent years,has received widespread attention and application.The deep learning model has the ability to learn features layer by layer.It learns the deep layer features of the original data through a multi-layer non-linear network structure,thereby effectively improving the ability to recognize the target.The main innovative work of this thesis is as follows:(1)Aiming at the one-sidedness and limitations of the single feature extraction method,this thesis proposes a target recognition method by jointly classifying three complementary features based on multi-task compressive sensing(Mt CS).The principal component analysis(PCA)features,elliptical Fourier descriptors(EFDs)and azimuthal sensitivity image(ASI)are extracted or constructed to describe the intensity distribution,target shape and electromagnetic characteristics of the original SAR images,respectively.The three features describe the original SAR image from different aspects thus their joint use can provide more discrimination for distinguishing different classes of targets.Afterwards,the three features are jointly classified based on Mt CS,which can not only properly represent individual tasks but also exploit their inner correlations to improve the performance of target recognition.(2)In order to improve the efficiency of target recognition and the antiinterference to noise,a multi-level method based on multi-level domain scattering images(DSIs)is proposed.First,multi-level DSIs are generated by selecting a certain number of main scattering points in the original image.They describes the relative positions and amplitudes of the dominant scattering centers and can be used to better discriminate different classes of targets.In addition,the background noises,which are usually with low intensities,can be effectively eliminated with high efficiency.Second,a prescreening of the multi-level DSIs is performed for individual test samples.Only those DSIs with higher reliability levels are used for the following target classification.Finally,joint sparse representation is employed to jointly classify the selected DSIs.The experimental results show that this method can not only improve the classification efficiency,but also effectively eliminate the background noise with low intensity.(3)In order to solve the problem that the decisions with low reliability will impair the fused performance to some extent,the reliability analysis is introduced into decision fusion for SAR target recognition.First,the reliability level of the decision is designed according to the fusion strategy,and only those with high reliability levels are used in the final decision fusion.In order to verify the validity of the reliability analysis,the proposed strategy is applied to multi-feature decision fusion and multi-classifier decision respectively.The experimental results show that this method has a higher recognition rate than the method that directly performs decision fusion.(4)Aiming at the problems that the small sample data set of MSTAR,which is difficult to meet the requirements of deep learning training,and the deep redundancy defects inherent in the residual network structure,a network model based on dense residual topological structure is proposed.The model connects the residual blocks through dense connection,and conveys the SAR image features learned layer by layer to the subsequent layers,avoiding the problem of deep redundancy of the residual network.At the same time,this model can further alleviate the gradient disappearance and gradient explosion problems of the network.In order to solve the problem that the number of samples of the MSTAR data set is small,the methods of mirroring,panning,scaling,rotation,and noise are used to expand the number of SAR image samples,which ensures the input required for training the network and avoids overfitting.