Robust And Efficient Cooperative Learning For SAR Imagery Of Multiple Features

Synthetic aperture radar(SAR)technology is an important microwave observation method,which has the characteristics of all-weather,long-distance and high resolution,and has important application value in the fields of aerospace,ground detection,battlefield detection and weapon guidance.With the development of SAR imaging technology,radar can achieve higher and higher resolution,which means that it can provide more and more detailed features.Whether the features are obvious directly affects the subsequent target detection and recognition.The traditional matching filter imaging method cannot break through the limitation of Nyquist sampling theorem,so it has low data utilization and produces high side lobe which affects the original feature.Emerging compressed-sensing algorithms conduct imaging based on sparse features of data.On the one hand,its robustness and efficiency need to be improved,on the other hand,it is difficult to protect other features in the process.There are few multi-feature enhanced imaging algorithms,and they are simply cascaded to deal with multiple features,which will cause error propagation.Therefore,it is of great significance to study a robust and efficient imaging method,which can coordinate the optimization of multiple features.This paper aims at the problem of multiple features enhancement in high-resolution SAR imaging,and starting from the data of the original two-dimensional wavenumber domain or the phase history domain after pre-processing.The method is used to design reasonable regularization to represent different prior information.At the same time,the dual decomposition and augmented Lagrange method are used to improve the efficiency and robustness of the solution,and the multi-feature collaborative learning is realized through the variable splitting.The main research contents are as follows.1.Research on SAR imaging and feature enhancement using compressed sampling data.Firstly,observed signal model is derived for stationary scene,ground moving target and air target,and the linear least squares problem of the inverse imagery is proposed.By improving real-valued alternating direction multiplier method(ADMM)algorithm,the complex-valued ADMM framework is introduced for multi-task learning.In this framework,as long as the introduction of the preset regularization,it can achieve higher resolution of SAR imaging and strengthen the corresponding features of the scene or target.In the experiment,both simulated and measured data are applied to validate the proposed framework.Compared with the conventional compressed sensing methods,the superiority of the proposed framework can be evaluated.2.Research on the sparse feature enhancement of point/block scattering distribution in high resolution SAR imaging scene.The higher resolution SAR imaging can provide more available features,such as point sparse features of isolated scattering point,block sparse features of gathering the strong scattering point and morphological sparsity features of different blocks.In this part,l1 norm,l2 norm block and the l2 norm with prior morphological information are introduced to enhance those three sparse features.The effectiveness of the three sparse feature enhancement methods is verified by experiments.3.Research on enhancing the directional texture feature of ground object in SAR imaging.To enhance the direction and texture feature of observation scene(such as the cropland),prior of directional text feature is introduced.First of all,the proposed algorithm builds the structure tensor,which is decomposed into singular values.Next,singular value weights divergence operator and builds the regular term.And then,the amplitude and phase separation method are applied to enhance the directional and sparse features and smooth texture structure.Finally,airborne spotlight data is used to verify the validity of the method.4.Research on the method of motion error joint auto-calibration in airborne SAR imaging.Because the airborne SAR platform will be affected by the airflow and deviate from the ideal track,coherent azimuth phase error(APE)and non-systematic range cell migration(Ns RCM)appearing in the data,which causes defocused imaging results.The Entropy function measure focusing performance,but it is not convex.Aiming at this problem,this part firstly uses the reasonable cost criterion to transform the entropy expression into a convex cost function and deduces the regular term corresponding to the focused feature.Then,the framework of this thesis is used to solve the sparse and entropy norms joint minimization problem,based on two-dimensional wavenumber-domain data.In the algorithm,APE and Ns RCM are synchronously updated to realize the joint error auto-calibration driven by data.Finally,airborne measured data(0.15m)is used to verify the effectiveness of the proposed method.