SAR图像目标特征提取与分类方法研究

【作者】 计科峰；

【导师】 郁文贤；

【作者基本信息】 中国人民解放军国防科学技术大学 ， 信息与通信工程， 2003， 博士

【摘要】 目前，SAR已经成为一种不可或缺的军事侦察手段。面对不断增长的SAR图像数据收集能力，如何对这些图像进行自动或半自动快速、准确地解译已经越来越引起人们的关注和重视。SAR ATR是自动或半自动SAR图像解译研究的一个重要方面。利用特征基于模型的SAR ATR系统代表着SAR ATR的发展趋势。本文系统研究了利用特征基于模型SAR ATR系统中，SAR图像目标特征提取方法和分类方法。 峰值特征是SAR图像目标识别的重要特征，为了由SAR图像快速、精确地提取目标峰值特征，论文首先研究了SAR图像目标峰值特征提取方法，提出了一种子像素级精度的SAR图像目标峰值特征自动提取方法，其对目标峰值位置的估计精度可以达到子像素级。为了在SAR成像参数确定的情况下，尽可能增强目标峰值特征，论文研究了峰值特征增强的SAR目标成像方法。在W.Clem Karl等工作的基础上，导出了求解峰值特征增强SAR成像优化问题的准牛顿迭代方法，建立了SAR成像稀疏投影矩阵T_S，通过用T_S代替原始SAR成像投影矩阵了，有效提高了峰值特征增强SAR成像方法的计算效率。为了提高基于特征匹配的SAR ATR系统的分类效率，论文进一步研究了SAR图像目标方位角估计方法，提出了一种利用峰值特征基于线性回归的SAR目标方位角估计方法，该方法除了具有计算速度快、估计精度较高的特点之外，还能在估计方位角的同时，给出该估计的置信区间，从而能更好地满足利用特征基于模型SAR ATR系统的需要。 属性散射中心包含了更丰富的可用于目标分类识别的特征，但同时由于特征参数的维数更高，因此相应的特征提取方法更复杂。为了将属性散射中心特征用于SAR目标分类，第三章研究了SAR图像目标属性散射中心特征提取方法，提出了RD-AML-CLEAN SAR图像目标属性散射中心特征提取方法，该方法可由输入SAR图像快速、自动地提取目标的属性散射中心特征。 在研究特征提取方法基础上，论文第四章研究了利用特征基于模型的SAR目标分类方法。利用特征基于模型的SAR目标分类方法，通过计算提取特征矢量和预测特征矢量之间的似然函数达到目标分类的目的。为了计算该似然函数，需要利用提取特征矢量和预测特征矢量之间的对应关系。本章以基于属性散射中心特征的分类为例，深入研究了多—多对应和1—1对应特征似然函数的计算，通过将求解二分图最佳匹配的算法用于寻找特征之间的最优1—1对应关系，有效提高了1—1对应特征似然函数的计算效率，分析了1—1对应和多—多对应特征似然函数之间的关系，给出了两种次优的1—1对应特征似然函数计算方法。 最后，论文第五章在前面各章的基础上，设计了一个利用特征基于模型的SAR目标分类仿真实验系统。基于该系统，通过大量MSTAR SAR图像数据的分类实验， 国防科学技术大学研究生院学位论文验证了本文特征提取方法以及分类方法的有效性，系统深入的分析了多个因素对SAR目标分类性能的影响。更多还原

【Abstract】 Recently, SAR is evolving to become an indispensable reconnaissance tool for military purposes. The collection capacity for SAR images is growing rapidly, and along with that growth is the expanding need for automated or semi-automated exploitation of SAR images accurately and efficiently. SAR ATR is an important aspect of automatic or semi-automatic SAR image interpretation. Model-based SAR ATR system that uses feature is the trend for SAR ATR. Methods of feature extraction and classification of target in model-based SAR ATR using feature are studied systemically in this paper.Peak feature is very important for SAR ATR. In order to extract target’s peak from SAR image rapidly and accurately, the method of target’s peak feature extraction is firstly studied in this dissertation. And a method of target’s peak automatic extraction in sub-pixel accuracy is proposed, it can estimate the position of peak in sub-pixel accuracy. For enhancing target’s peak feature with SAR imaging parameters given, the peak-enhanced SAR imaging method is studied. Based on work of W. Clem Karl etc., the quasi-Newton iteration method for solving the optimization problem of SAR imaging is derived, the sparse projection matrix Tsof SAR imaging is constructed, by replacing the original SAR imaging projection matrix T with TS, the computation efficiency of peak-enhanced SAR imaging is improved greatly. In order to improve the efficiency of classification based on feature matching, the method of azimuth estimation from SAR image is studied. A method of target’s azimuth estimation from SAR image using peak feature based on linear regression is proposed, besides goodish estimation accuracy and high computation efficiency, it can also provide the confidence interval of the estimation, which can meet the need of model-based SAR ATR system that uses feature very well.Attributed scattering center comprises more features for SAR ATR, but because of its high dimensionality, the method of attributed scattering center extraction is more complicated. In order to put attributed scattering center ’into SAR ATR use, the method of attributed scattering center extraction is studied in chapter 3. And the RD(Region Decoupled)-AML(Approximate Maximum Likelihood)-CLEAN method is presented for extracting attributed scattering center from SAR image, it can extract scattering center from SAR image quickly and automatically.Based on the method of feature extraction, the method of model-based SAR target classification using feature is studied. The model-based SAR target classifier that uses feature accomplishes classification by computing the likelihood function between extractedand predicted features. Computation the likelihood function requires using the correspondences between extracted and predicted features. Taking attributed scattering center-based classification as example, the computation of feature likelihood function under many-many and 1-1 correspondence are studied, by using the algorithm of bipartite graph perfect matching to find the optimal 1-1 correspondence, the computation efficiency is improved greatly, the relations of likelihood function between 1-1 and many-many correspondence are analyzed, and two sub-optimal methods of calculating the likelihood function of 1-1 correspondence are presented.Finally, based on preceding chapters, an experiment system of model-based SAR target classification using feature is devised in chapter 5, based on this system, through large number of experiments using MSTAR SAR image, the validities of methods of feature extraction and classification are verified, effects of several factors on performance of SAR target classification are analyzed thoroughly and systemically.更多还原