Group Sparse Representation-Based SAR Image Despeckling,Target Detection And Recognition

Synthetic Aperture Radar(SAR)is a general method for generation high-resolution radar maps from low-resolution aperture data,which is extensively used to image objects on the surface of the Earth.SAR plays an important role in military and civil field for situation awareness under all weather conditions,day or night.As a significant source of ground information,SAR images play an important role in providing powerful support to commands and decisions.Because of the coherent imaging mechanism of SAR system,SAR images is different from optical images,which makes it difficult to understand.The group sparse representation(GSR)theory gives more efficient representation of the structure and target information of SAR image,which provides another solution for SAR image speckle noise suppression,target detection and target recognition.In this thesis,we explores the group sparsity of SAR image with high resolution.By analyzing the measurements of the characteristic parameters of the ground at a deeper level,we enforce the GSR theory to solve the SAR image despeckling,target detection and target recognition problems,which is significant for the widespread use of SAR technology in military and civil field.The main research content of this thesis are listed as follows:1.In order to preserve the structural information and suppress speckle noise simultaneously,the GSR theory is utilized to exploit the local similarity of SAR image.By analyzing the properties of multiplicative model of speckle noise,a novel dictionary learning algorithm based GSR(GSRDL)for SAR image despeckling is proposed in this thesis.In order to overcome the influence of speckle noise,a mean filter is used to obtain the approximation of the filtered clean SAR image.Utilizing the local similarity of the filtered SAR image,we establish the objective function based on GSR model and mean filter.The optimal dictionary and group sparse representation coefficients are calculated by solving this objective function.In this way,the structure information with local similarity is reconstructed and the speckle noise is reduced simultaneously.Experimental results on SAR image despeckling manifest that the proposed GSR-DL algorithm achieves better performance than other state-of-the-art despeckling algorithms.2.In order to better describe the structural information and the details of SAR image,the learned dictionary obtained by GSRDL is used to reconstruct the low-frequency structural information and the prespecified dictionary is used to recover the high-frequency detail information.In this way,we proposed an over-complete dictionary design algorithm based on GSR(OD-GSR)for SAR image IV despeckling.The learned dictionary shows better performance in structural information preservation and speckle noise suppression.The prespecified dictionary we selected mainly consists of two parts which are the orthonormal basis of band-limited wavelets and a tight frame of band-limited shearlets.These two basis are used to reconstruct the high-frequency information of SAR image.Finally,the images recovered by learned dictionary and prespecified dictionary are merged together as the result.The experimental results on real SAR images demonstrate that the proposed OD-GSR algorithm can achieve more effective speckle reduction as well as image detail preservation.3.Because of the coherent imaging mechanism of SAR system,interesting targets in SAR images always appears as clusters which consist of clustered independent strong scatterers with the same size of targets.Utilizing these properties of targets,a novel target detection algorithm based on the ideal scattering center model and GSR is proposed.Firstly,we establish the GSR model of SAR target based on the ideal scattering center model.Secondly,dynamic group sparse algorithm is utilized to reconstruct the clustered scattering centers and suppress the discrete scattering centers.Finally,we select the targets which has the same size of interesting target.The experimental results show that the proposed target detection algorithm can achieve better target detection performance and suppress the occurrence of the false alarm.4.Since the SAR target often appears as a set of scattering center and it is sensitive to imaging orientation,the geometric features of targets are not stable.We proposed two-dimensional Euler-principal component analysis(e-2DPCA)to extract features of SAR targets.2DPCA which provide more accurate covariance matrix approximation and the extracted features preserve the structural information of SAR targets.And on this basis,we intended to extract the nonlinear correlation between pixels by combining with Euler kernel.In this way,more stable features are extracted wich can reduce the influence of outliers.Finally,the classification of SAR targets are realized by using K nearest neighbor classifier.The experimental results show that the proposed e-2DPCA algorithm gives a more efficient feature extraction for SAR images.Especially,the proposed feature extraction algorithm can guarantee better accurate recognition rate(ARR)under strong noise interference.5.In order to represent the targets with multiple features,we explore the correlation between different monogenic features and proposed a target recognition algorithm based on joint GSR and residual weighted sum.The weighted value is defined by the structural similarity of sparse representation vector.Firstly,the monogenic signal model is used to obtain the over-complete feature dictionaries over different scales.Based on these structure similarity of the coefficients,two strategies are proposed to improve the ARR of SAR target.The proposed algorithm enforces dynamic group sparse representation algorithm for joint group sparse representation between the same feature dictionary over different scales.As for the information fusion between different components,we propose a novel weighting strategy using the structural similarity of sparse representation vector.Finally,the results is calculated by searching the minimum value of residual over different subdictionaries constructed by different kinds of training images.Experimental results demonstrate the outstanding performance of the proposed algorithm.