基于稀疏特征的SAR图像处理与应用研究

发布时间：2018-08-27 16:03

【摘要】：合成孔径雷达(SAR)是一种工作在微波波段的高分辨率、主动式相干成像雷达,其成像具有全天候、远距离、能适应恶劣环境等优点,在对地观测中发挥着重要的作用。随着分辨率、成像方式等技术的进步,SAR图像数据量的急剧膨胀给数据传输和存储带来极大的压力,SAR相干成像的特性带来的相干斑噪声给SAR图像的解译和应用带来困难。从SAR图像的稀疏特性来看,稀疏表示理论在SAR图像处理与应用领域具有广阔的前景。开展基于稀疏特征的SAR图像处理与应用方法研究,可以提高SAR图像处理和解译的水平,推广SAR图像在民用领域和军事领域的应用。本文从SAR图像的稀疏特征出发,基于字典学习和稀疏模型,对SAR图像压缩、SAR图像相干斑抑制、SAR图像目标分类和SAR图像与红外图像融合进行了研究。本文主要工作总结如下:第二章研究内容:在SAR图像信号盲稀疏度条件下,现有重构算法中固定阈值的选择限制了SAR图像压缩后重构精度和重构速度的提高。鉴于此,本文提出一种改进的正交匹配追踪算法。该算法通过非线性下降的阈值快速选择原子,自动调节候选集原子个数,以便每一次迭代时更加精确地估计真正的支撑集,利用正则化过程实现了支撑集的第二次筛选,实现了盲稀疏度信号的精确重构。本文将算法应用于SAR图像压缩,提出了一种基于稀疏表示的SAR图像压缩方法。实验结果表明,本文方法能在压缩SAR图像的同时抑制相干斑。第三章研究内容:提出了一种基于自蛇扩散和稀疏表示的Contourlet域SAR图像相干斑抑制方法。对SAR图像Contourlet变换分解后的低频子带采用自蛇扩散处理,并将滤波处理后的系数作为SAR图像低频子带在Contourlet域的局部均值估计;基于稀疏优化模型,利用改进的正交匹配追踪算法求解高频子带的稀疏系数,重构得到高频子带系数;对滤波后的所有子带系数进行Contourlet逆变换,实现SAR图像的相干斑抑制。实验结果表明,该方法能在抑制相干斑噪声的同时较好地保留图像的边缘信息。第四章研究内容:针对SAR图像的目标分类,本文提出了三种方法。(1)基于多子分类器的Ad Boost分类方法。该方法首先提取训练样本的2D-LDA特征和G2DPCA特征,采用SVM方法分别训练弱分类器,采用Ada Boost.M2算法将弱分类器提升为强分类器完成SAR图像的目标分类。实验结果表明,该方法在识别率和识别时间上优于采用单一子分类器的Ada Boost算法。(2)基于EMACH和稀疏表示的分类方法。该方法针对SAR图像训练样本,采用EMACH算法训练模板,提取模板的G2DPCA特征构造过完备字典,利用改进的正交匹配追踪算法求解测试样本的G2DPCA特征在过完备字典下的稀疏系数,根据系数的能量特征判断SAR图像目标的类别。实验结果表明,该方法在分类速度方面明显优于同类别的其他分类方法。(3)提出了一种基于级联字典和稀疏表示的分类方法。该方法利用SAR图像训练样本库,按类别生成多个字典,构成级联结构,测试样本经过级联字典,按需依次求解稀疏系数,利用重构误差和投票机制完成SAR图像的目标分类。实验结果表明,该方法明显提高了SAR图像的目标分类速度。第五章研究内容:针对SAR图像与红外图像融合,本文提出了两种融合方法。(1)基于自适应权值的曲波域SAR图像与红外图像融合方法。该方法引入模糊理论的分析思想,对源图像Curvelet变换后的不同频率子带系数采用不同的权值进行处理,通过自适应加权策略实现SAR图像与红外图像融合。实验结果表明,该方法真实可靠,能够有效提高融合质量。(2)基于稀疏表示的NSCT域SAR图像与红外图像融合方法。该方法首先将源图像进行NSCT分解,形成不同频带的系数;采用区域能量融合方法融合低频子带;构造过完备字典,求解不同高频子带在字典下的稀疏系数,采用能量最大原则选取稀疏系数并重构高频子带;通过NSCT逆变换将不同频带系数融合,实现SAR图像与红外图像融合。实验结果表明,本文融合方法在视觉效果和客观指标评价方面,都行之有效。
[Abstract]:Synthetic Aperture Radar (SAR) is a high-resolution, active coherent imaging radar operating in microwave band. Its imaging has the advantages of all-weather, long-distance, and can adapt to harsh environment. It plays an important role in earth observation. With the development of resolution, imaging methods and other technologies, the rapid expansion of SAR image data to data transmission. Sparse representation theory has broad prospects in the field of SAR image processing and application from the point of view of sparse characteristics of SAR images. In order to improve the level of SAR image processing and interpretation and popularize the application of SAR image in civil and military fields, this paper studies SAR image compression, speckle suppression, target classification and fusion of SAR image and infrared image based on the sparse features of SAR image, dictionary learning and sparse model. The work is summarized as follows: In the second chapter, under the condition of blind sparsity of SAR image signal, the selection of fixed threshold limits the improvement of reconstruction precision and reconstruction speed after SAR image compression. Selecting atoms and automatically adjusting the number of atoms in the candidate set are used to estimate the true support set more accurately in each iteration. The second selection of the support set is realized by the regularization process, and the blind sparseness signal is reconstructed accurately. Experimental results show that the proposed method can suppress speckles while compressing SAR images. Chapter 3: A speckle suppression method for SAR images in Contourlet domain based on Self-Snake diffusion and sparse representation is proposed. Based on the sparse optimization model, the sparse coefficients of the high-frequency subband are solved by the improved orthogonal matching pursuit algorithm, and the high-frequency subband coefficients are reconstructed. All the filtered subband coefficients are inversely transformed by the Contourlet transform to achieve speckle suppression. The experimental results show that this method can suppress speckle noise and preserve image edge information. Chapter 4: Aiming at target classification of SAR images, three methods are proposed. (1) Ad Boost classification method based on multi-sub-classifier. The weak classifier is trained by SVM method, and the weak classifier is upgraded to strong classifier by Ada Boost. M2 algorithm. The experimental results show that this method is superior to Ada Boost algorithm with single Subclassifier in recognition rate and recognition time. (2) Classification method based on EMACH and sparse representation. The training samples are trained by EMACH algorithm, and the G2DPCA features of the template are extracted to construct an over-complete dictionary. The sparse coefficients of G2DPCA features in the over-complete dictionary of the test samples are solved by the improved orthogonal matching pursuit algorithm, and the classification of SAR image targets is judged according to the energy features of the coefficients. (3) A classification method based on cascaded dictionary and sparse representation is proposed, which uses the SAR image training sample library to generate multiple dictionaries according to the category, and forms a cascaded structure. The test samples are cascaded dictionaries, and the sparse coefficients are solved in turn according to the need. The reconstruction error and the input are used. The experimental results show that the proposed method improves the classification speed of SAR image. Chapter 5: Aiming at the fusion of SAR image and infrared image, two fusion methods are proposed. (1) Curvilinear SAR image and infrared image fusion method based on adaptive weights. According to the analysis idea of fuzzy theory, different weights are used to process the sub-band coefficients of different frequencies of the source image after Curvelet transform, and the fusion of SAR image and infrared image is realized by adaptive weighting strategy. The experimental results show that this method is reliable and can effectively improve the fusion quality. (2) SAR image and red image in NSCT domain based on sparse representation First, the source image is decomposed by NSCT to form the coefficients of different frequency bands; then the low-frequency sub-bands are fused by regional energy fusion method; then the sparse coefficients of different high-frequency sub-bands are solved by constructing an over-complete dictionary; then the sparse coefficients are selected and the high-frequency sub-bands are reconstructed by using the principle of maximum energy; and the inversion of NSCT is performed. Different band coefficients are fused to realize the fusion of SAR image and infrared image. The experimental results show that the fusion method is effective in visual effect and objective evaluation.
【学位授予单位】：南京航空航天大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN957.52

【相似文献】