【作者】 陶勇；

【导师】 胡卫东；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2010， 博士

【摘要】 SAR作为一种有效的信息获取手段,被广泛地应用于军事侦察和地理遥感等多个领域。随着SAR数据收集能力的不断提高,如何快速、有效地从SAR图像中提取出感兴趣目标类型的目标识别,就成为制约战场情报生成和态势感知能力的重要环节。目前SAR图像目标识别面临的最大技术难题来自于目标特性的多样性和所处环境的复杂性,有限的SAR图像信息不足以刻画目标特性的变化,还需要充分地挖掘和利用其它辅助信息。针对此问题,本文将领域知识引入到识别中,分别从知识表示、结构模型描述以及知识应用三个方面探讨和研究知识辅助的SAR图像目标识别。首先,从知识的基本含义出发,介绍了在SAR图像目标识别中需要用到的有关目标内部构成关系的结构模型知识和有关目标和外部环境间因果关系的上下文知识。基于散射中心频率依赖性和方位依赖性中蕴含的结构模型信息,简要地说明了两种结构模型知识的表示和匹配方法。对于经验性的上下文知识,从证据和知识不确定性的表示、不确定性的传递和结论不确定性的合成几个方面讨论了各种不确定性知识的表示和处理方法。然后,在结构模型知识研究方面,首先从散射中心的频率依赖性角度研究了高分辨雷达目标一维结构模型的描述,即模型参数的估计方法。在分析了子空间型方法对GTD模型位置参数估计的有效性后,提出了一种基于正交投影的模型参数估计方法。该方法依据GTD模型中散射中心位置和类型参数的低耦合,利用信号子空间的平移不变结构估计散射中心位置,根据频率依赖项的正交投影系数对类型参数进行鉴别,使整个估计过程一次完成,具有较低的计算复杂性。分析表明,该方法具有较高的估计精度和良好的推广性能,从而为SAR图像目标二维结构模型的研究奠定了基础。其后,同时考虑散射中心的频率依赖性和方位依赖性,对基于属性散射中心模型的SAR图像目标二维结构模型的描述展开研究。在总结图像域模型参数提取基本流程的基础上,分析了图像分割对参数估计性能的重要影响。为指导散射中心区域分割,将频率域中的散射模型转换到图像域,对属性散射中心的分布特性进行了充分地分析。根据属性散射中心模型中局域式和展布式散射中心方位特性的差异,提出了一种类型判断-图像分割-参数估计的模型参数顺序估计方法。为了提高参数估计性能,分析了基于lk范数的SAR复图像域正则化方法实现超分辨的内在机理,并针对该方法在不同强度散射点条件下分辨率提高不一致的问题,提出了一种基于变参数的改进正则化方法。这些研究完备了二维结构模型知识的应用基础。最后,对结构模型知识和上下文知识辅助的SAR图像目标识别方法进行研究。根据结构模型知识的层次化结构,提出了一种知识辅助的序贯处理流程,并分析了特定结构知识在SAR图像目标识别中的优势。基于全局结构知识预测得到的主散射中心结构,提出了一种构造近似旋转平移不变特征的匹配识别方法。在阐述了上下文知识在SAR图像目标识别中重要性后,提出了一种上下文知识辅助的目标识别模型,并以可信度方法为例对知识处理的过程进行了具体说明。更多还原

【Abstract】 As an effective sensor for acquiring information, synthetic aperture radar (SAR) is now widely used in military reconnaissance as well as geophysical remote sensing applications. Along with the increasing capacity of SAR collection, automated or semi-automated SAR target recognition, which allows quickly and efficiently identification of interested objects in SAR image, has become a crucial part of battlefield intelligence gathering and situation assessment. At present the principal challenge of SAR image target recognition is root in the variability of SAR signatures of targets and complexity of background environment and the limitied information in SAR observation is badly unsufficient in coping with this problem. That is to say, apart from SAR observation, other aided information such as expert knowledge should be introduced. Focusing on exploitation of expert knowledge for SAR target recognition, three key techniques such as knowledge representation, structural model description and knowledge application are studied in this dissertation.Starting from the basic meaning of knowledge, this paper firstly introduces the structural model knowledge and the contextural knowledge. The two types of knowledge respectively describe the internal structure of target and the causality between the target and the external environment. They should be very useful in SAR target recognition. Based on the structural model information of the frequency dependence and the aspect dependence in scattering center model, two kinds of structural model knowledge representation are briefly established. Aiming at the uncertainty in the experiential contextual knowledge, this paper discusses several kinds of representation and processing method of uncertain knowledge. The uncertainty representation of evidence and knowledge, uncertainty propagation and uncertainty combination of conclusions are subsequently elaborated.Secondly, in the aspect of structural model knowledge, according to the frequency dependence of scattering center in geometrical theory of diffraction (GTD) model, one dimensional structural model of high resolution radar target is studied, namely model parameter estimation method. After analyzing the validity of position parameter estimation of GTD model using subspace-type technique, an orthogonal projection-based method for parameter estimation is presented. According to weak coupling between scattering center position and type parameter in GTD model, this method firstly utilizes the shift-invariance characteristerics of signal subspace to estimate the position of scattering center, and then determine the type of scattering center in terms of the orthogonal projection of frequency-dependent term. The sequential estimation schema largely reduces computational complexity. Moreover theoristic analysis and simulated experiment illustrate that the method can simultaneously achieve accurate estimation and well bandwidth-generalized performance, which lays a foundation for the research of two dimensional structural model.Subsequently, based on the frenquency dependency and the aspect dependency in attributed scattering center model, two dimensional structural model of target in SAR image is studied. After reviewing the main flow of complex imagery-domain model parameters extraction, the effects of image segmentation is analyzed. Aiming at the above issue, we transform the scattering center model from the frequency-aspect domain into the image domain and analyze the distribution characteristics of attributed scattering centers fully, which provides a useful guidance for the segmentation of SAR image and type selection of attributed scattering centers. According to different aspect-dependence characteristics of localized and extended scattering centers in attributed scattering center model, a three steps of type selection, image segmentation and parameter estimation is proposed. Moreover, in order to improve the performance of parameter estimation, the inherent principles of lk norm regularization-based super-resolution in SAR complex image domain are studied, and a modified lk norm regularization method is presented. Owing to varying parameters, the proposed method can mitigate the problem of inconsistent resolution for scattering centers with different amplitudes. The above researches have completed the apllication basis of structural model knowledge.Finally, the application of the structural model knowledge and the contextual knowledge to SAR image target recognition is investigated. According to the hierarchy of the structural model knowledge, a sequential knowledge-aided processing flow is proposed. Therein the specific structural knowledge describes the local relation of stuctures, and its merit in SAR image target recognition is analyzed. Moreover, we employ the global structural knowledge to construct the main scattering structure and put forward an approximately translational and rotational invariant feature based attributed point matching algorithm for target recognition. On the other hand, the importance of the contextual knowledge in SAR image target recognition is stated. A model of contextual knowledge-aided target recognition is presented, and the confidence method as an application example is utilized to illustrate the detailed process.更多还原