【作者】 尹奎英；

【导师】 金林； 刘宏伟；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2011， 博士

【摘要】 合成孔径雷达(SAR)地面目标识别在战场感知中发挥着重要作用,是地面情报的主要来源之一,能够为战争决策提供强有力的支持。本论文主要对SAR地面目标识别技术进行了深入的研究,从两个方面展开:其一为SAR图像预处理,其二为SAR目标特征提取和SAR目标识别。论文内容可概括为如下六部分:第一部分,提出了一种复合的SAR图像去噪算法。该方法首先利用信号的小波相邻尺度相关性将信号和噪声分离,然后根据SAR图像近似瑞利分布的特点把SAR图像转换为近高斯分布,再分别利用复数扩散震动滤波器对SAR图像部分进行增强,用各项异性扩散方程对含噪部分进行去噪,最后用小波平稳变换对图像进行重构。实验结果表明,与传统的SAR图像去噪算法相比,新算法无论在边缘增强还是在噪声去除能力方面均有显著提高。第二部分,提出了一种基于CUDA分段自回归模型的SAR图像插值算法。由于硬件条件和SAR成像的限制,获取的SAR图像分辨率不高,不利于图像的观察和分析,通常的最近邻插值法、双线性插值法、双三次插值法、立方卷积插值法等对原始图像进行放大,一定程度上改善了图像的视觉效果,但是这些算法插值的精细度不够高,对图像的视觉效果改善度不够大。基于分段二维自回归模型(PAR)的图像插值算法,通过使用非线性优化的方法来同时估计模型参数和缺失像素。这种插值方法能够产生具有很好视觉效果的高分辨率图像,但是该方法的计算成本非常高。针对该问题,提出了一个基于CUDA的并行实现方法,它先将一幅图像根据两次插值约束分成多个9?9的局部窗口,对每个局部窗口进行基于分段自回归模型的图像插值:首先,启动一个CUDA线程利用自回归插值算法来对图像进行第一轮插值,而自回归模型的参数则利用高精度的共轭梯度算法进行估计;然后,在第二轮插值中,利用第一轮插值得到的自回归模型参数作为共轭梯度算法的初始值进行第二轮的参数估计,从而降低了第二轮插值时间。实验结果表明,与基于传统的CPU的插值算法相比,基于GPU的并行算法在实现插值一个2592?1944的图像时,时间缩短为原来的1/110,并且图像尺寸越大,加速越快。文中的算法能最佳的利用自回归模型的局部性和CUDA的并行性,在得到高质量插值图像的同时计算速度大幅提高。第三部分,提出了一种融合SAR目标轮廓和阴影轮廓的目标识别算法。在进行SAR自动识别时,大多数算法都使用目标内部结构特征,如灰度值,峰值,中心距等特征,而很少使用SAR图像轮廓信息,更很少使用SAR图像阴影信息。事实上,SAR图像轮廓反映了SAR目标的局部空间结构信息,如果分割得当,能够分割出准确,精细的轮廓信息,而轮廓信息可以作为一种非常稳健的识别特征的。因此我们首先提出了一种基于控制标记符的SAR图像分割算法,得到SAR图像目标轮廓和阴影轮廓,然后融合这两种轮廓进行SAR目标识别。基于MSTAR的实验结果验证了本算法的有效性。实验结果证明,目标轮廓和阴影轮廓的结合,除反映本身包含的局部空间结构信息外,还能反应SAR目标的高度信息,较单一轮廓特征,是一种更为稳健的特征。第四部分,提出了一种基于多模分布的SAR图像分割算法。SAR图像目标,背景,阴影的不同成像机理使得这三部分具有不同的统计特性。本文分析SAR图像三部分的统计性质,并对其分别建立统计模型,并给出了一种基于这三种模型组合的多模分布的SAR图像分割算法,对于目标分割和阴影分割分别采用不同的预处理方法,分别提出了快速Otsu分割算法分割目标,以及基于背景均值保留的冲击滤波算法分割阴影。分割结果表明,这种基于多模分布的SAR图像分割算法与传统的基于单模分割算法相比,能最佳的利用各部分统计特性的差异,得到准确的分割结果。第五部分,提出了一种SAR目标姿态角的估计算法。SAR成像对目标方位角非常敏感,当SAR与目标的相对位置发生变化时,目标的散射中心也会发生变化,导致不同的方位角下的目标有明显的区别。因此,SAR目标姿态角的估计是SAR目标识别的一个重要步骤,在SAR目标分类和识别中,精确的方位角估计可以减小目标匹配数和检测误差。我们在前面目标分割的基础上,分析SAR图像随方位角变化的不变特征,提出了一种基于清晰双边的随机Hough变换法,基于MSTAR的实测数据实验结果表明,这种算法估算SAR目标姿态角具有估计精度高,计算时间短的优点。第六部分,提出了一种基于纹理特征的SAR目标识别算法,同时针对SAR目标变体提出了一种基于局部纹理特征的SAR目标变体识别算法。由于真实世界中使用的测试数据不可能与训练数据完全一致,也就是同一目标存在多种不同变体,这是SAR目标识别的难点,也是影响SAR目标识别率和SAR目标识别推广性的主要原因之一。我们提出了一种针对变体的识别算法,利用变体与原目标局部纹理之间的相似性进行识别。首先,提出了一种基于清晰边缘的SAR图像配准算法,然后使用结合Gabor变换,LBP和空间区域直方图的纹理特征来描述SAR图像,最后用基于大特征的直方图序列的匹配做识别。由于采用了基于配准的纹理来描述SAR图像,因而能有效描述SAR目标;用局部直方图匹配来进行识别,比用基于全局特征的算法推广性更好。基于MSTAR S2的实验结果证明了本算法的有效性。更多还原

【Abstract】 Synthetic Aperture Radar (SAR) ground target recognition plays an important role in battlefield awareness. As one of major source of ground information, it can provide powerful support to war decision. This dissertation addresses issues of SAR ground target recognition technology. The work of this dissertation mainly focuses on two aspects: The first one is SAR image pre-processing, and the second one are SAR target characteristics extraction and SAR target recognition. The main content of this dissertation is summarized as follows.The first part of this dissertation presents a composite adaptive enhancing and denoising algorithm for SAR images. The SAR images are differentiated from speckle noise via scale space correlation. Because SAR image can be described by approximate rayleigh’s distribution, complex diffusion coupled shock filter is used to enhance the signal differentiated from SAR image, and the anisotropic diffusion equation is used to denoise the speckle in SAR image. At last, the SAR image is reconstructed by stationary wavelet transform. Compared with traditional speckle removal algorithms, this new algorithm has better performance in terms of edge preserving and denoising.Based on CUDA, the second part of this dissertation proposes a parallel algorithm to speed up the SAR image interpolation algorithm based on the piecewise autoregressive model. Due to hardware condition and SAR imaging limitation, the resolution of obtained SAR image is not high enough for observation and analysis. Common image interpolation algorithms, such as nearest neighbor, bilinear interpolation, bicubic interpolation, cubic convolution interpolation, can be used to enlarge the original image. These algorithms can help to improve the visual effect of image in some degree. But limited by the fineness of algorithms, the improvement to visual effect of image is not high enough. The piece-wise 2D autoregressive (PAR) model interpolation algorithm uses non-linear optimization method to evaluate model parameters and missing pixels simultaneously. Because of good adaptability to image local pixel structure, this algorithm can get a high image reproduction quality. But the computational cost of this algorithm is also very high. A parallel algorithm based on CUDA is proposed to speed up the image interpolation algorithm based on the piecewise autoregressive model. An image is first divided into multiple 9?9 small local windows. For each local window, first, a CUDA thread is used to interpolate the local window using the autoregressive interpolation algorithm whose parameters are found by the high accuracy gradient descent algorithm; in the second interpolation round, the gradient descent algorithm uses the estimated parameters of the autoregressive model found in the first interpolation as the initial values to decrease the computation time in the second interpolation round. Numerical simulation indicates that this GPU based parallel algorithm can interpolate a 2592?1944 image within 1/110 of the time used by a CPU based algorithm. Moreover, the computation time saved decreases with the image size. Since this algorithm can parallel process the pixel interpolation on GPU, compared with traditional serial algorithm that runs on CPU, this parallel implementation can make better use of the local property of the piecewise autoregressive model and parallel property of CUDA, eventually achieving an interpolation image of high quality in a low computation time.The third part of this dissertation, a SAR image segmentation algorithm based on multi-mode distribution is proposed. In SAR image, the generations of image, target and background clutter are different, which give rise to their different statistical characteristics. In this part, an analysis of the different imaging characteristics of the three components is made and statistical models of them are developed, on the basis of which, a multi-mode distribution SAR image segmentation algorithm based on the combination of the three models is proposed. For the target and shadows segmentation, different pre-processing methods are adopted. A fast one-dimensional Otsu algorithm is proposed to segment target from background clutter and a SAR image enhancement algorithm based on the background mean-preserving shock filter to segment shadow from background clutter. Experimental results show that, compared with traditional segmentation algorithms based on the single-mode distributions, this multi-mode distribution segmentation algorithm can make the best use of the differences of statistical characteristics of the three components, with high segmentation accuracy and a better practicability.The fourth part of this dissertation gives an SAR target recognition algorithm based on fusion of target contour and shadow contour. When performing SAR automatic target recognition, the internal structure characteristics of target are most frequently used, such as gray value, peak value and center distance. SAR image contour information is rarely used in target recognition and SAR image shadow information is even less used. Actually, SAR image contour can reflect local spatial structure information of SAR target. If SAR image is segmented appropriately, correct and fine contour information can be acquired. Contour information is a kind of stable recognition characteristic. Therefore, to acquire target contour and shadow contour, we propose an SAR image segmentation algorithm based on marker-controlled. Then, fusion of these two kinds of contour is used to perform SAR target recognition. The effectiveness of this algorithm is verified through experimental results on MSTAR data. Experimental results show that, besides local spatial structure information, fusion of target contour and shadow contour also contains height information of SAR target. Compared with these two features used independently, fusion of contour is a kind of more stable characteristic.The fifth part of this dissertion presents an estimation algorithm for SAR target attitude angle. SAR imaging is very sensitive to target azimuth angle. When relative position between SAR and target is changed, scattering center of target will chage. This causes same target varies significantly under different azimuth angle. Therefore, SAR target attitude angle estimation is an important step of SAR target recognition. During SAR target classification and recognition, precise azimuth angle estimation can help to reduce target matching number and detection error. Based on the target segmentation mentioned above, we analyze the unchanged characteristics of SAR image under different azimuth angle and propose a Hough transformation algorithm based on clear double edge. Experimental result based on MSTAR measured data show that, using this algorithm to estimate SAR target attitude angle can achieve high precision and short computation time.Based on texture characteristic, the sixth part of this dissertation gives a SAR image recognition algorithm. Aiming at SAR target variant, a SAR target variant recognition algorithm based on local texture characteristic is also included in this part. Since one target usually has multiple variants in practice, the measured data in the real world is quite different from the training data, which cause the SAR automatic target recognition (ATR) difficult and becomes one of the major factors affecting SAR target recognition rate. Therefore this paper presents a SAR ATR algorithm aiming at the target variants. The new algorithm uses the local texture similarity between the variant and the original target for recognition. Firstly, a SAR image registration algorithm based on clear edges is proposed. Then, the texture characteristic obtained by a combined use of the Gabor transform, LBP and spatial domain histogram is employed to describe the SAR image. At last, histogram sequence matching based on the large characteristic is used to perform the recognition. Since texture characteristic based on registration is used to describe the SAR image, the SAR target can be described effectively. Furthermore, local histogram matching is used to perform the recognition; therefore the new algorithm has better generalization property than algorithms based on the global characteristic. The effectiveness of the proposed algorithm is verified by experimental results on MSTAR S2.更多还原