【作者】 王世晞；

【导师】 沈振康；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2008， 博士

【摘要】 本文以发展自动和半自动的SAR图像目标分类系统为目的,围绕成像侦察情报支援作战应用,结合我国雷达成像卫星及航天器与地面应用系统建设及发展需求,以SAR图像目标解译应用为背景,在大量国内外高分辨率SAR实测数据的支持下,深入研究了SAR图像去斑技术、SAR图像分割技术、面向目标分类的SAR图像目标切片特征提取技术。论文主要研究内容包括:第二章研究了SAR图像的相干斑抑制问题。相干斑抑制是SAR图像处理中最基础同时也是最重要的问题之一,其核心问题在于如何在充分抑制相干斑的同时保持图像的点、线、边缘等结构特征。论文首先提出了一种具有结构保持特性的MRF模型——SPMRF,并给出了模型的参数估计方法,该模型可以根据图像的局部特征自适应调整权重参数,既能描述匀质区域又能描述结构特征,为贝叶斯估计提供了准确的先验信息,从而使基于SPMRF的SAR图像贝叶斯去斑取得了较好的去斑效果;然后,借鉴自适应窗口滤波思想,提出了MRF邻域的自适应调整方法,弥补了简单MRF模型无法保持结构特征的缺陷。通过对图像局部区域匀质性的判断,自适应调整模型的邻域结构,在匀质区域使用较大邻域以充分抑制相干斑,而在含结构特征区域使用较小邻域,并筛选出与中心像素最有可能源自具有相同后向散射特性的邻域点参与计算,以保持结构特征。以此为基础的AN-MMRF去斑在充分抑制相干斑的同时较好地保持了图像结构特征;最后,在已有的基于HMT和基于HMRF模型的隐状态估计方法上,将两者结合,提出了基于HMT-HMRF模型的SAR图像小波系数隐状态估计法,该方法充分利用了小波系数尺度间和尺度内的相关性,提高了隐状态估计的准确性,为利用贝叶斯估计削弱噪声主导的小波系数并保持信号主导的小波系数奠定了基础。以此为核心的SAR图像小波去斑同样取得了较好的去斑效果。第三章研究了SAR图像区域分割技术。为了从大幅未知的SAR场景图像有效提取目标ROI切片并分离切片中的目标区域,论文给出了三种分割算法,包括基于最大类间方差准则的SAR图像分割算法、基于分形特征组的SAR图像分割算法和基于多分辨率分析的SAR图像分割算法。首先,在现有的二维最大类间方差法分割算法的基础上,分析了叠加乘性噪声的二维直方图特点,提出新的适用于乘性噪声的直方图区域划分方法。同时,提出新的阈值选取准则。基于改进的二维直方图划分方法和新的阈值选取准则,论文提出了基于最大类间方差准则的SAR图像分割算法;其次,针对SAR图像纹理的特征,利用分形理论来计算待分割像素局部图像数据的分形维数和间隙度特征,以衡量该局部图像数据的起伏特性。论文基于这两类特征构建分形特征矢量,并结合二项式距离判决函数,实现SAR图像分割处理;最后,为了消除相干斑噪声对高分辨率SAR图像分割的影响,论文给出了基于多分辨率分析的SAR图像分割算法。该算法在对待分割图像数据进行多尺度分层处理的基础上,对多尺度数据建立MAR模型,并计算多分辨率对数似然比统计量,来实现对SAR图像的分割。第四章研究了面向目标分类的SAR图像目标特征提取技术。立足于构建自动和半自动的SAR图像目标分类系统,深入研究了面向SAR图像目标分类的目标切片特征提取问题。首先,在广泛文献调研的基础上,综述了SAR图像目标方位角估计技术;在此基础上,提出一种主导边界与最小外接矩形联合的SAR目标方位角估计方法。该方法充分考虑了基于主导边界和最小外接矩形方法的优缺点,取长补短,大大提高了目标方位角估计的精度;然后,致力于构建半自动的SAR图像目标分类系统,研究了面向人机交互的图像目标几何特征提取与分析问题,提取了几种直观的、有效的、便于判读员理解的目标几何特征;最后,以自动目标分类系统的实时性指标为主要考虑依据,提出了一种快速的SAR目标识别方法。该方法采用基于Hebb学习规则的主分量分析(PCA)进行特征提取,使用多层感知器神经网络(MLP NN)进行目标分类。实现了自动快速的目标分类。更多还原

【Abstract】 Aiming at developing automatic and semi-automatic systems of SAR image classification, considering imaging reconnaissance applications and the demand of building and developing radar imaging satellites, spacecraft and ground application systems, with the background of SAR image target interpretation applications and the support of abundant high-resolution SAR data, this thesis comprehensively studies the techniques of SAR image despeckling, SAR image segmentation, feature extraction from SAR chips for target classification. The research works of this thesis are as follows:SAR image despeckling is developed in chapter 2. Despeckling is a basic and important subject of SAR image processing, which focuses on suppressing the speckle while preserving the structural features such as point, line and edge. First, a structure-preserving MRF model, the SPMRF model, is proposed and the parameter estimation method is given. This model can adaptively adjust the weights according to the local features in image. It can be used to describe both homogeneous regions and structural features. It provides accurate prior knowledge and leads to good despeckling performance. Then, based on the idea of adaptive window, a method of adaptively adjusting the MRF neighborhood is presented, which solves the problem that simple MRF model can’t preserve the structural feature. The neighborhood is adjusted via determining the homogeneity of local region. Larger neighborhood is adopted in homogeneous regions to suppress speckle. Smaller neighborhood is adopted in regions with structural features to preserve the structural feature by selecting the pixels which are most likely to have the same backscattering properties as the center pixel. In this way, the AN-MMRF despeckling method preserves the structural feature while suppressing speckle. Finally, combining the HMT-based and HMRF-based hidden state estimation methods, a HMT-HMRF-model-based method for SAR image wavelet coefficients is proposed, which utilizes the correlation within and between scales of wavelet coefficients and improves the accuracy of estimation. The corresponding wavelet despeckling method achieves good performance.SAR image segmentation is discussed in chapter 3. In order to extract target ROI chips from large-scene SAR images and separating the target regions, three segmentation methods are studied, namely, segmentation based on maximum-between-class-variance, segmentation based on fractal features and segmentation based on multi-resolution analysis. First, based on the existing maximum-between-class-variance segmentation, 2D histogram is analyzed on image with multiplicative noise and a new method of segmenting histogram with multiplicative noise is proposed. The rule for determining threshold is also provided. Based on the improved 2D histogram segmentation and new thresholding method, SAR image segmentation based on maximum-between-class–variance is presented. Second, considering the textural feature of SAR image, fractal theory is used to compute the fractal dimension and spacing feature to measure the fluctuating properties of local data. Based on the two features and quadratic distance function, SAR image segmentation is implemented. Finally, in order to remove the influence of speckle on high-resolution SAR image segmentation, segmentation method based on multi-resolution analysis is proposed. In this method, multi-resolution processing is implemented on images, and then MAR model are built on multi-scale data. Multi-resolution log-likelihood ratio is computed to achieving SAR image segmentation.Target feature extraction for classification is presented in chapter 4. Established in constructing an automatic and semi-automatic SAR image target classification system, the target chip extraction for SAR target classification is deeply studied. First, the techniques to estimate SAR target orientation angle are reviewed, and a method jointing dominant boundaries and minimum outer rectangles is proposed to estimate the SAR target orientation angle. The method, considering the advantages and disadvantages of that only uses dominant boundaries or outer rectangles, can significantly improve the estimation accuracy of target orientation angle. Then, established in constructing a semi-automatic SAR target classification system, the extraction of image geometric features is studied, and several intuitive and effective target geometric features are extracted. Finally, a fast SAR target recognition method is proposed. The method uses Principle Component Analysis (PCA) based on learning rules to extract target features, and Multiple Levels Perception Neural Network (MLPNN) to classify the target. The automatic and fast target classification is thus finished.更多还原