【作者】 赵炳爱；

【导师】 范晓虹； 苏辉； 邱志明； 胡新生；

【作者基本信息】 哈尔滨工程大学 ， 管理科学与工程， 2003， 博士

【摘要】 雷达目标识别系统在战场上的应用对现代战争的发展有重大的意义，增强目标识别系统分类算法对动态多变环境的自适应性，提高识别系统的智能化，是雷达目标识别领域乃至整个自动目标识别(ATR)领域的一个关键性问题。合成孔径雷达(SAR)技术不断发展，星载与机载SAR成像数量极大，在SAR图像分析中应用ATR技术，加强对高分辨率雷达图像目标的识别分类具有重要价值。 小波分析理论对信号的分解具有良好的局域化特性，运用小波分析理论可以在多尺度下分析目标特征，符合人脑对信息加工的基本特点，成为雷达目标信号处理强有力的数学工具。 SAR图像的斑点噪声对识别率影响较大，在目标识别前需把斑点从图像中去除掉。传统的滤波技术在一定条件下很难满足滤波的高性能要求，而利用基于小波分析的滤波方法在保持目标边缘特征方面具有很大的优势。面对大量的滤波算法，根据不同图像类型和任务要求，选用不同滤波方法的研究急待深入。本文对采用空间域常用的滤波算法、时间频率域相结合的滤波算法滤除SAR图像噪声的结果进行了比较。分别选用增强Lee滤波算法、无偏GMAP算法、小波系数压缩滤波算法、小波变换与维纳滤波相结合、小波变换与空间滤波相结合的方法对四幅不同的SAR图像进行滤波，指出了各滤波算法对不同图像处理的优缺点，给出了在不同应用环境下，选择使用不同滤波器的意见。本文根据SAR图像小波分析中各级细节子图方差间存在的近似线性关系及经验的阈值比例系数改进了常用的小波滤波方法，在保证精度的基础上提高了运算速度。 二维小波变换可使图像目标特性在不同分辨率下显露出来，所得细节图像和近似图像可在多分辨下反映分布目标的纹理等特征，具有更强的类别可分性，对SAR图像分析十分有意义。 传统二进小波变换缺少模式识别所需的移不变特征，需要确定适合于模式识别的具有移不变特性的二维小波变换。本文对传统二维小波变换与过完全小波变换进行了区别和比较；并对常规的二维小波变换、无下采样的小波变换与多孔算法小波变换提取纹理能量特征进行了研究与对照比较。本文首次利用图像灰度均值与无下采样小波变换细节图像纹理能量组成特征矢量，使目标间纹理特征的差异比Seisuke Fukuda等方法所得更明显。在比较利用传统小波变换与过完全小波算法提取SAR图像纹理能量特征的主要区别过程中，发现本文采用过完全小波变换提取的纹理特征方便宜用，完全可以作为SAR图像面目标的统计特征。 人工神经网络是模式识别的重要工具，本文分别采用BP神经网络、径向基函数(RBF)神经网络、自组织特征映射神经网络对SAR图像面目标进行了分析，选用灰度值、均值、小波纹理特征等不同的特征作为输入矢量，得到了高的分类精度。本文首次将过完全小波分解提取的新的纹理特征 (O认找rF特征)矢量与BP、RBF神经网络相结合对目标进行分类识别，分类结果表明，OWAI下特征矢量与RBF神经网络相结合处理结果更好。 传统K均值算法、模糊C一均值聚类算法及自组织特征映射神经网络等图像分类方法，在自动分类过程中以像素灰度为分析对象，这是运用这些方法分类运算速度慢的根本原因。本文将直方图作为一个重要因素引入到算汀、中，使迭代对象从每个像素的灰值变为图像存在的灰度级，运算量降至原算法的幻(M只N)，由于所占内存量相应变小，运算速度进一步提高，可在很短}}寸间内完成对图像的分割。 本文针对三彩色SAR图像或胶片扫描SAR图像，采用了基于饱和度经验距离的K均值分类算法进行分析，并与针对图像灰值进行分析的结果加以比较。通过比较发现，采用基于饱和度经验距离的K均值算法分类的结果体现了各个区域同类目标间的一致性，视觉效果更好。 在图像目标识别中，基于统计的和基于人工智能的模式识别方法都无法适]fl」二日标识别的各个方面。对于强噪声背景下的机场SAR图像目标识别问题，木文分别利用二值和灰值形态梯度方法进行分析，该方法在保持目标形状和突出强噪声背景下重要目标信J息、方面优于常规算法。本文将形态统讨-滤波器一与方向小波分解结合起来提取SA尺图像弱边缘，在图像中很好地分割出「!标与背景，为海岸线提取提供了有效途径。 本文首次采用累积概率分布及K一S距离对海域SAR图像目标进行检测，少夯提出综合分析整幅图像及目标区域图像确定闭值的方法，利用此法对图像进行二值化的效果优于常规方法，为海域目标的分类研究做好了准备。更多还原

【Abstract】 Radar target recognition systems are very important in modern wars. The adaptation of targets classing algorithms to variable environment, the intelligent standards of the systems, are of critical in the field of radar target recognition, even in the field of auto target recognition (ATR). With the development of SAR (synthetic aperture radar) technology, a large amount of images are made by space-borne and airborne SAR, it’s important to use ATR technology in SAR images analysis and advance the level of target recognition of high-resolution radar.With the theory of wavelet analysis, the decomposition of the signal is localized. The characteristic of the signal can be analyzed in different levels, conforming to the basic methods of human brains’ work style. Wavelet analysis is becoming a useful tool in radar signal processing now.Speckle in SAR images will affect the precision of recognition, should be removed before classing. While traditional filtering algorithms are not efficient to filter the speckle, combining wavelet analysis with other algorithms have the priority in keeping the edge of the objections.Though there are many methods in the field of speckle filtering, it lacks deep study of methods selection with different images and in different ATR missions. In this paper, both speckles filtering in space and wavelet based methods is used to reduce the speckle. Analysis is done to compare the result of filtering. The paper uses enhanced Lee filter, unbiased GMAP algorithm, shrinking of wavelet coefficients algorithm, wavelet analysis with Wiener filter, wavelet analysis with space filter to filter the speckle in four different SAR images, and gives a suggestion for the choice of right filter in different situation. With experimental threshold coefficient and approximate scale relation between sub-images of wavelet detail coefficients, an improved algorithm is used to get a faster speed with more precise threshold.With 2D-wavelet transform, the characters of image objects can be found in different resolution, texture features can be extracted form detail and approximate images of wavelet transformation. It’s very useful in SAR images analysis. Dyadic wavelet transform is not shift-invariant, which is always requested in pattern recognition, so new 2D-wavelet transform method is needed. The paper compared the difference of ordinary 2D-wavelet transform and over-complete 2D-wavelet analysis, also studied the difference in characteristic extraction of ordinary, over-complete wavelet transform and ’a trous’ method. In this article,vectors are composed of mean gray level of image and texture features in detail images of 2D-wavelet transform without down-sampling. Distances between each two objects’ vectors are much larger than that of Seisuke Fukuda’s. In the process of comparison, it is found that the new vector used by this paper is suited to be taken as statistic feature of SAR area objects.Artificial neural network (ANN) is always used in image classing. This thesis used BP network, RBF network and SOFM network to analyze SAR area objects, with gray level, average and wavelet analysis based features as the inputs. The precision of the result is high. In both BP and RBF network, the inputs are made up by OWATF (Over-complete wavelet analysis based texture feature) vectors, with small training samples, the result of the later one is much better than that of the previous one.In image classing process, the speed of the algorithms is an important factor. K-means algorithm, FCM, SOFM network, all deal image pixels as the object, which is the main reason of long working time. By introducing histogram to the algorithms, the computing quantity of the new algorithms is about K/(M X N) to the originals’. Accordingly, the amount of needed memory is also reduced. The classing procedure is sped up, and the result can be got in very short time. In the paper, saturation based experienced distance is used in K-Mean algorithm. With this method, color SAR images and scanned film SAR images are analyzed.更多还原