【作者】 朱磊；

【导师】 水鹏朗；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2014， 博士

【摘要】 合成孔径雷达(synthetic aperture radar, SAR)由于具有全天时与全天候等优点，因此在环境信息获取与目标检测等众多领域发挥着重要作用。然而与其它相干成像系统一样，由SAR系统形成的SAR图像中散布着大量乘性的相干斑噪声。大量随机分布的相干斑不仅使得SAR图像的视觉效果大为下降，而且极大的制约了SAR图像特征提取、目标跟踪等解译处理技术的可靠性与有效性。因此，SAR图像相干斑的抑制研究具有重要的理论与现实意义。本论文主要围绕SAR图像等效视数(equivalent number of looks, ENL)估计与相干斑抑制方法展开研究，论文的主要内容可概括为以下四个部分：基于边缘强度映射(edge strength map, ESM)，提出了一种SAR图像ENL参量的自适应非监督估计方法。其一，借助各向异性高斯核(anisotropic Gaussian kernel,AGK)平行窗估计比率ESM；其二，对SAR图像进行分块处理并利用一种简单的非监督估计方法分别估计各分块图像在比率ESM下的区域阈值；其三，利用估计的区域阈值与阈值化ESM操作分别实现对各分块图像的区域划分；其四，借助比率ESM两次约束下的大尺度不规则局域窗估计所有非边缘区像素的局部ENL；其五，借助直方图统计方法从众多局部ENL估计值中提取全局ENL估计。实验表明了该估计方法的有效性与稳定性。以相干斑在小波域的一种加性转换噪声模型为基础，提出了一种基于非下采样小波包变换(undecimated wavelet packet transform, UWPT)的SAR图像变换域抑斑方法。其一，利用UWPT变换实现对SAR图像的多层非下采样子带分解；其二，借助自蛇扩散抑制低通子带中残留的相干斑噪声，并将抑斑后的低通子带系数视为原始SAR图像在小波域的一种局部均值估计；其三，利用小波域局部均值估计自适应调节改进型L1-L2联合优化，并实现对所有高频子带的自适应软阈值滤波去噪；其四，通过重构抑斑后的各子带完成对SAR图像的相干斑抑制处理。实验表明该方法在SAR图像的相干斑抑制与边缘保护方面均取得了较好的效果。提出了三种SAR图像空域抑斑方法。(1)基于区域划分的空域抑斑方法。该方法先借助具有方向性的高斯-伽马平行窗估计比率ESM，然后通过阈值化ESM操作实现对SAR图像边缘与非边缘区域的划分，进而分别利用改进的Kuan滤波与Sigma滤波平滑SAR图像不同区域的相干斑噪声。另外，为强化相干斑抑制效果，迭代滤波策略与变尺度局域窗被采用。(2)基于改进Frost滤波的空域抑斑方法。该方法以传统Frost滤波采用的负指数衰减型加权滤波模型为基础，通过将SAR图像多种局部统计参量结合作为联合衰减因子，形成与SAR图像区域分布特性相适应的负指数型加权系数，同时采取两次滤波策略，先由预滤波削弱SAR图像相干斑噪声并估计获得更精准的局部统计参量，然后借助精细局部统计参量再对原始SAR图像实施精细滤波。(3)基于迭代方向滤波的空域抑斑方法。该方法先借助高斯-伽马平行窗估计出的比率ESM与方向信息自适应地控制AGK生成沿ESM方向分布的具有各向异性支撑区域的局域窗，然后将SAR图像多种局部统计参量联合作为衰减因子，形成与SAR图像区域分布特性相适应的负指数衰减型加权系数，进而将负指数衰减型加权系数同局域窗带方向的各向异性支撑区域结合形成局域加权的方向滤波，最后对SAR图像迭代地实施方向滤波，实现带边缘保护的相干斑抑制。实验证明了上述三种空域抑斑方法在保护SAR图像边缘的同时，可有效抑制相干斑噪声。提出了两种SAR图像各向异性扩散抑斑方法。(1)基于平均曲率运动(meancurvature motion，MCM)的改进DPAD抑斑方法。该方法通过将MCM耦合到经典的各向异性扩散抑斑算法DPAD中，并利用Kuan滤波系数生成耦合函数，用于控制MCM的扩散强度，形成了一种可有效抑制边缘区域噪声与同质区块效应的各向异性扩散抑斑方法。(2)带方向约束的各向异性扩散抑斑方法。通过将局部方向比率估计的方向约束同改进的Frost滤波结合形成带方向约束的扩散函数，然后将带方向约束的扩散函数嵌入与MCM扩散耦合的各向异性扩散框架，从而形成一种带方向约束的各向异性扩散抑斑方法。实验证明了上述两种各向异性扩散抑斑方法在保护SAR图像边缘的同时，可有效抑制边缘区域的相干斑噪声与同质区的块效应现象。更多还原

【Abstract】 Synthetic aperture radar (SAR) plays an important role in many fields, such asgathering information from the Earth’s environment and detecting target，due to itspower ability to implement daytime and nighttime measurements under all weatherconditions. Like other coherent imaging systems, SAR images acquired from a SARsystem inevitably suffer from the multiplicative random noise called speckle. Thepresence of speckle seriously degrades the visual quality of SAR images and limits theeffectiveness for subsequent interpretation processing technologies, such as featureextracting technology and object tracking technology.Therefore, speckle suppressionhas important significance to improve the imaging quality of SAR images and theinterpretation processing effect.This thesis mainly considers the estimating method of equivalent number of looks(ENL) and the despeckling methods for SAR images. The main contributions of thethesis are summarized as the following four parts:Based on edge strength map (ESM), an unsupervised estimation method for theENL in SAR images is given. First, the ESM is produced by ratio operation ofanisotropic Gaussian kernel (AGK) parallel windows. Second, SAR image is dividedinto several image blocks by simple image partition processing, and the local ESMthresholds for image blocks are obtained by an effective unsupervised estimationmethod. Third, SAR image is divided into homogeneous regions and edge regions byshrinkage ESM. Fourth, each local ENL of all pixels in homogeneous regions isestimated by large irregular window under twice restrictions based on ESM. Fifth,global ENL of SAR image is acquired by histogram statistical method for all local ENLvalues. The experimental results show that the proposed method is effective and stabile.Based on an additive transform noise mode of speckle noise in SAR image, atransform-domain despeckling method is proposed by using undecimated waveletpacket transform(UWPT). First, a SAR image is decomposed into multiple subbands bymulti-level UWPT. Second, the lowpass subband is filtered by self-snake diffusion, andthe filtered lowpass subband was regarded as the local mean of the original SAR imagein wavelet domain. Third, adaptive and shrinkage soft-thresholding filter was applied tothe rest subbands by improved L1-L2optimization based on the local mean. Fourth, thedespeckled image was recovered from the all of filtered subbands by the inverse UWPT.The experimental results show the proposed method has good performance in reducing speckle and preserving the edge of SAR images.Three spatial-domain despeckling methods for SAR images are studied.(1)Spatial-domain despeckling method based on region subdivision is given. First,Gaussian-Gamma-shaped bi-windows with different orientations are applied to estimatethe ESM by virtue of ratio operations. Second, the different regions in SAR image areobtained by the threshold-processing ESM. Third, two improved filters, the Kuan filterand the Sigma filter, are used to smooth speckle in homogeneous and edge regions,respectively. In order to strength the ability to reduce speckle in SAR images, iterationfiltering strategy and variable scale local window are adopted in the proposed method.(2) Spatial-domain despeckling method based on improved Frost filtering is proposed.The new method is based on the negative-exponential and weighted filtering model intranditional Frost filter. The weighting coefficients of the new method are obtained bythe decay factor using several local statistics and are adaptive to the characteristics ofregional distribution of SAR image. Meanwhile, two-stage filtering strategy is used.First, the pre-filtering is executed for reducing speckle in SAR image and estimatingfine local statistics. Second, the fine filtering for original SAR image is carried outbased on the fine local statistics obtained by the pre-filtering.(3) Spatial-domaindespeckling method based on iterative direction filtering is given. First, the ratio ESMand direction information are estimated by Gaussian-Gamma-shaped bi-windows, andanisotropic support domain along the ESM direction is obtained by using the ESM anddirection information to self-adaptively control the AGK in rectangular local window.Second, the decay factor is obtained by combining several local statistics, and thenegative-exponential weighting coefficients produced by the decay factor are adaptiveto the characteristics of regional distribution of SAR image. Third, direction filtering isformed by combining the negative-exponential weighting coefficients and the localwindow with anisotropic support domain and different directions. Lastly, reducingdespeckle in SAR image with edge protection can be realized by iterative operation ofdirection filtering. The experimental results show the proposed three spatial-domaindespeckling methods can effectively reduce speckle while preserving the edges in SARimages.Two anisotropic diffusion despeckling methods for SAR images are conducted.(1)Improved DPAD despeckling method based on mean curvature motion (MCM) isproposed. By embedding the MCM into the classical DPAD and using the coefficientsof Kuan filter to control diffusion intensity of the MCM, the improved DPADdespeckling method is developed. The improved DPAD can effectively smooth speckle near edges and reduce blocking artifacts in homogeneous regions.(2) Despecklingmethod via direction-constrained anisotropic diffusion is introduced. Adirection-constrained anisotropic function is developed by combining the improvedFrost filter and the direction constraint obtained by the local directional ratios (LDRs).By the embedding anisotropic function with direction constraint into the anisotropicdiffusion framework coupled with the MCM, the new despeckling method viadirection-constrained anisotropic diffusion is proposed. The experimental results showthe proposed two anisotropic diffusion despeckling methods can effectively reducespeckle near edges and blocking artifacts in homogeneous regions.更多还原