【作者】 杨萌；

【导师】 张弓；

【作者基本信息】 南京航空航天大学 ， 通信与信息系统， 2012， 博士

【摘要】 合成孔径雷达(SAR)作为主动寻的探测设备，在对地观测中发挥着重要作用，是陆海情报的主要来源之一，能够为指挥决策提供强有力的支持。由于SAR相干成像特性，SAR图像中固有的相干斑和地物电磁散射特征是SAR图像的解译和应用的分析基础。从SAR图像目标的特点和表现形式，以及信号稀疏性分析的发展来看，稀疏表示技术在SAR图像处理的理论和应用领域具有广阔的应用前景。开展基于稀疏表示的SAR图像处理关键技术研究，对于提高SAR图像处理和解译水平，推广SAR图像在民用及军事领域的应用具有深远的意义。本文利用稀疏表示理论对SAR图像相干斑抑制、目标检测和变化检测进行了深入的研究。本文主要工作总结如下：第二章研究内容：提出了K-OLS超完备字典的构造算法，并应用于SAR图像相干斑抑制。K-OLS超完备字典的构造算法利用基于K均值聚类的向量量化原理和正交最小二乘算法(OLS)，通过分步优化字典原子和变换系数自适应构造了超完备字典。SAR图像相干斑抑制算法，首先利用获得的K-OLS超完备字典将图像局部信息投影到高维空间中，实现图像的稀疏表示；其次运用正则化方法建立多目标优化模型；最后通过对优化问题的求解重建SAR图像场景分辨单元的平均强度，从而实现了SAR图像的相干斑抑制过程。实验结果表明，算法对相干斑噪声有很好的抑制效果。第三章研究内容：提出了SAR图像稀疏表示的多元稀疏优化模型，并将其应用于SAR图像相干斑抑制。针对SAR图像点、线、面的结构特征，本文运用具有点奇异性的小波、具有线奇异性的剪切波、具有面奇异性的K-OLS字典，通过正则化方法建立了多元稀疏优化模型。SAR图像相干斑抑制算法通过对多元稀疏优化模型的求解，重建SAR图像场景分辨单元的平均强度，实现了SAR图像的相干斑抑制。实验结果表明，该算法对SAR图像相干斑噪声具有很好的抑制效果，且相比于K-OLS方法，该算法具有增强图像纹理细节特征的优点。第四章研究内容：针对SAR图像中感兴趣目标的稀疏性，(1)提出了基于超完备二维离散傅里叶变换字典的SAR图像自动目标检测算法。基于超完备字典学习的稀疏表示建立在过完备基础上，具有较强的数据稀疏性和建模稳健性。该算法首先通过构造超完备二维离散傅里叶变换字典将SAR图像数据投影到高维空间,实现了图像局部特征的稀疏表示；然后利用随机矩阵获得稀疏域局部特征的压缩采样，并对多组采样数据运用聚类算法并行处理；最后通过符号检验法，实现了对目标像素与背景像素的分类。实验表明，算法对硬目标具有较好的检测效果；(2)从反问题的角度，提出了基于理想点散射中心模型的无监督SAR图像目标检测算法。算法首先通过散射中心模型构造超完备字典将图像目标信息投影到频率-方位角二维空间中，实现图像的稀疏表示；其次运用随机矩阵得到了数据压缩域特征子空间；最后利用聚类算法和概率投票方法进行像素分类，实现SAR图像目标的检测。实验结果表明，算法不仅能够很好的检测出SAR图像的目标，而且对相干斑噪声具有很好的鲁棒性。第五章研究内容：针对SAR图像变化检测的鲁棒性问题，(1)提出了基于改进K-SVD的SAR图像变化检测算法。算法首先通过构造超完备字典将图像局部信息投影到高维空间中，实现图像的稀疏表示；其次运用随机矩阵得到了数据在高维空间中的低维特征子空间；最后利用模糊聚类算法进行无监督聚类，实现SAR图像变化区域信息的重构。实验结果表明，算法不仅能够很好的检测出图像的区域变化，而且对噪声具有很好的鲁棒性；(2)针对图像的二维信号特性，给出了二维信号的压缩感知框架，并将其应用于SAR图像变化检测问题。提出了基于散射中心模型的SAR图像变化检测算法。算法首先通过点散射中心模型获得SAR图像局部特征的稀疏表示；其次运用二维压缩感知理论进行压缩采样；最后利用模糊聚类算法进行无监督聚类，实现SAR图像变化区域信息的重构。实验结果表明，算法具有较好的检测性能和噪声鲁棒性。更多还原

【Abstract】 As an active sensing, Synthetic Aperture Radar (SAR) plays an important role in situationawareness in sea, and land environments. As one of major sources of ground information, it canprovide powerful support to command and decision. This dissertation addresses issues ofdespeckling and detection in SAR image. As SAR is a coherent imaging system, the informationfor interpretation is carried by the average intensity or Radar Cross Section (RCS) at each speckledpixel. It is obvious that sparse representation has good prospects for the application in SAR imageprocessing. It is important to improve level of research and interpretation by carrying out keytechnology research of the sparse representation in SAR image processing. It is also important forSAR image processing in civil and military applications. It is a complex process. There are someremained theoretical problems to be systematically solved. In this paper, SAR image specklesuppression, target detection and change detection based on the sparse representation theory arediscussed in depth. The main research work and contributions of this paper are shown as follows:Chapter two: A de-speckling algorithm for SAR images using adaptive over-complete learneddictionary is proposed. A K-OLS algorithm for designing overcomplete dictionary for sparserepresentation is proposed. The vector quantization based on K means algorithm andorthogonal least square algorithm (OLS) are used and a practical optimization strategy basedon an iterative loop is used to design a redundant dictionary. A de-speckling algorithm forSAR images using K-OLS algorithm is proposed. Firstly, SAR image is projected into a highdimensional space using the learned dictionary and a sparse representation of SAR image isobtained. Secondly, model for multi-objective optimization problem is built by regulationmethod. Finally, the de-noising process is realized through solution of the multi-objectiveoptimization problem in which the mean backscatter power is reconstructed. The experimentalresults demonstrate that the proposed algorithm has good de-speckling capability.Chapter three: A new methodology for despeckling of SAR images using sparse optimizationmodel is proposed. The algorithm based on sparse representation via over-complete dictionaryhave a strong data sparseness and provide solid modeling assumptions for data sets. Firstly, asparse optimization model based on structural properties of SAR image is built by regulation.Secondly, a practical optimization strategy is used to design a redundancy dictionary. And then,a over-complete dictionary is constructed by employing a combined dictionary consisting of wavelets, shearlets and redundancy dictionary. Finally, the despeckling process is realizedthrough solution of the multi-objective optimization problem in which the mean backscatterpower is reconstructed. The experimental results demonstrate that the proposed algorithm hasgood de-speckling capability and advantages of enhancing image details.Chapter four: In connection with the sparseness problems of target in SAR images,(1) theautomatic target detection algorithm due to the inherent sparsity of target in SAR image isproposed. Firstly, a two-dimensional discrete cosine transform dictionary is constructed toproject the SAR image into a high dimensional space and a sparse representation set of imagelocal features is achieved. Secondly, random sampling matrix is used to do compressionsampling and mean shift algorithm is applied to handle multiple sets of sample data withparallel processing. Finally, the algorithm achieves the target pixels and background pixelsclassification using the sign test method. The experimental results demonstrate that theproposed algorithms have a good target segmentation results for hard target in SAR images;(2)from the point of view of the inverse problem, we introduce a new method for target detectionin SAR images using point scattering center model based on the target backscattercharacteristics. For this algorithm, the image is projected onto frequency-aspect space throughscattering center model, giving an adaptive sparse representation. Random matrix is taken asmeasurement matrix to realize generation of the feature space. And then, the final targetdetection is realized by clustering algorithm and probabilistic voting method, achieving thereconstruction of target regional information. The experimental results demonstrate that theproposed algorithms have a good target detection results and also have a good robustness onthe speckle noise.Chapter five: In connection with the robustness problems in change detection of SAR images,(1) we introduce a new method for change detection in remote sensing images using sparserepresentstion. For the algorithm, a large collection of image patches is projected onto highdimensional spaces through improved K-SVD dictionary, giving a sparse representation pereach image patch. Random matrix is taken as measurement matrix to realize feature spacedimension reduction. And then, the final change detection is realized by clustering the featurevector space using the Fuzzy Clustering algorithm, achieving the reconstruction of changeregional information. The experimental results demonstrate that the proposed algorithms havea good change detection results both in contour and region and also have a good robustness onthe noise;(2) we introduce a new framework for two-dimensional compressed sensing and a new method for change detection in remote sensing images using two-dimensionalcompressed sensing. For the change detection algorithm, a large collection of image patches isprojected onto high dimensional spaces through improved overcomplete dictionary, giving asparse representation per each image patch. Two random matries are taken as measurementmatrix to realize feature space dimension reduction. And then, the final change detection isrealized by clustering the feature vector space using the Fuzzy Clustering algorithm, achievingthe reconstruction of change regional information. The experimental results demonstrate thatthe proposed algorithms have a good change detection results and also have a good robustnesson the noise.更多还原