合成孔径雷达图像局部特征提取与应用研究

发布时间：2018-06-19 23:54

本文选题：SAR图像 + 局部特征　；参考：《国防科学技术大学》2016年博士论文

【摘要】：SAR图像特征提取是SAR图像信息提取与解译的基础工作,良好的特征能够完备地表征图像或者目标的核心信息,有利于后续的识别分类工作的开展,因而受到了国内外研究人员的关注,成为SAR图像解译的理论研究热点之一。在SAR图像解译的实际应用中,图像和目标的全局特征提取通常是非鲁棒的,存在一定随机变化和模糊。SAR图像局部特征提取研究是解决上述问题的一种思路,是对特征提取技术的进一步发展,也是SAR图像计算机解译技术的重要信息支撑。而目前SAR图像局部特征提取普遍沿用光学图像处理的方法,存在一定的不相适应的情况,因此根据SAR图像成像特性与噪声特性来研究其局部特征提取算法很有必要。通过研究与完善SAR图像局部特征理论与提取算法,克服两个不足:(1)传统的全局特征提取方法鲁棒性差;(2)光学图像局部特征提取方法对于SAR图像的适应性差,从而提高基于局部特征的SAR图像匹配与目标检测识别的精度与效率。本文围绕SAR图像局部特征提取的这一主线,从SAR图像目标的显著性特征、目标的散射中心点集特征和SAR图像局部不变特征三个方向入手,采用理论分析和实验验证相结合的研究方法,深入研究探讨:(1)SAR图像目标显著性区域、显著性特征提取与应用;(2)目标散射中心点集特征序贯匹配识别SAR图像车辆目标;(3)SAR图像局部不变特征提取新方法与应用。作为整个论文的理论基础,第二章主要论述了图像局部不变特征理论及典型的方法,SAR图像的成像概述、噪声特性和几种基本特征,针对本文研究重点,对SAR图像局部不变特征的研究基础、发展现状和主要方法做了阐述和总结。第三章在SAR图像噪声特性分析基础上,结合视觉显著性理论,根据SAR图像局部复杂度和自差异测度设计了一种SAR图像的显著性区域检测方法,与经典的视觉显著性算法进行显著性区域检测实验对比,以及与经典的CFAR方法进行SAR图像目标检测实验对比,均取得了良好的实验结果,表明该方法胜任在SAR图像上提取显著性区域,检测高价值地物目标,具有很好的应用性。第四章结合模式识别理论中的点模式匹配方法和SAR图像属性散射中心模型理论,提出了一种基于目标散射中心点集特征的序贯匹配方法。首先基于属性散射中心模型提取SAR图像域目标散射中心的特征,使用该散射中心点集位置分量和表征散射中心几何结构信息的属性散射中心特征频率影响因子作为匹配特征,依次序贯匹配,实现车辆目标的对比识别。实验结果与国外同类方法相比,准确性和识别率都显示出优势。第五章提出了一种SAR图像局部特征提取新方法。首先分析了SAR图像的像素梯度信息提取方法和常见的梯度算子,例如ROA算子、ROEWA算子和GR算子。然后介绍了Harris算子、Lo G算子及其适应SAR图像特性的改进—多尺度Harris算子。接下来介绍了图像局部二值模式特征和旋转不变局部二值模式特征,多尺度局部梯度比率直方图特征MLGRPH,并且实验验证了该特征在SAR图像上的旋转不变性能。在上述研究基础上,提出了基于MLGRPH特征的SAR图像局部不变特征提取新方法。实验验证环节,采用不同时相、不同波段、不同极化方式和不同视角成像的多组SAR数据对经典SIFT方法、SIFT-OCT方法和该新方法开展了实验分析与对比,结果表明该方法在性能上优于SIFT方法、SIFT-OCT方法,还可以进一步性能提升的潜力。第六章对本文的工作总结归纳,并对SAR图像局部特征提取方法的下一步研究工作进行了展望。
[Abstract]:The feature extraction of SAR image is the basic work of information extraction and interpretation of SAR images. Good features can fully characterize the core information of images or targets, which is beneficial to the follow-up recognition and classification work. Therefore, it has attracted the attention of researchers both at home and abroad. It has become one of the hot topics in the theoretical research of SAR image interpretation. The interpretation of SAR images is interpreted. In practical application, the global feature extraction of image and target is usually non robust. There is a certain random change and local feature extraction of fuzzy.SAR image. It is a way to solve the above problems. It is the further development of the feature extraction technology and an important information support for the SAR image computing machine interpretation technology. And the current SAR diagram It is necessary to study the local feature extraction algorithm based on the image characteristics and noise characteristics of SAR image, so it is necessary to study and improve the local feature theory and extraction algorithm of SAR image, so as to overcome two shortcomings: (1) traditional global special. The robustness of the extraction method is poor; (2) the local feature extraction method of the optical image is poor in the adaptability of the SAR image, thus improving the accuracy and efficiency of the SAR image matching based on the local feature and the target detection recognition. This paper focuses on the main line of the local feature extraction of the SAR image, from the saliency feature of the target of the SAR image and the scattering center of the target. Starting with three directions of point set feature and local invariant feature of SAR image, the research method combined with theoretical analysis and experimental verification is used to study and discuss: (1) significant feature extraction and application of SAR image target, (2) the feature sequential matching of target scattering center set recognition for SAR image vehicle target; (3) local SAR image is not local. A new method and application of variable feature extraction. As the theoretical basis of the whole thesis, the second chapter mainly discusses the image local invariant feature theory and typical methods, the overview of the image, the noise characteristics and several basic features of the SAR image. In this paper, the research foundation, the development status and the main methods for the local invariant features of the SAR image are discussed. The third chapter, based on the analysis of SAR image noise characteristics, combined with the visual significance theory, designed a significant regional detection method for SAR images based on the local complexity and self difference measure of SAR images, compared with the classic visual saliency algorithm for significant regional detection experiments, and with the classic CFAR. The method is compared with the SAR image target detection experiment, and good experimental results are obtained. It shows that the method is competent for extracting significant regions on SAR images and detecting high value objects. The fourth chapter combines the point pattern matching method and the SAR image attribute scattering center model theory in the pattern recognition theory. A sequential matching method based on the feature of the target scattering center point set is proposed. Firstly, the feature of the target scattering center of the SAR image domain is extracted based on the attribute scattering center model, and the characteristic frequency influence factor of the attribute scattering center of the scattering center is used as the matching feature. In the fifth chapter, a new method for extracting local features of SAR images is proposed. Firstly, the method of extracting the pixel gradient information from the SAR image and the common gradient operators, such as ROA operator, ROEWA calculation, are analyzed. Then, the Harris operator, the Lo G operator and the improvement of the Lo G operator and the improvement of the SAR image characteristics are introduced. Then the features of the local two value pattern and the rotation invariant local two value pattern, the multi-scale local gradient ratio histogram feature MLGRPH are introduced, and the experiment verifies that the feature is on the SAR image. On the basis of the above research, a new method for extracting local invariant features of SAR images based on MLGRPH features is proposed. Experimental verification links are carried out by using different phases, different bands, different polarization modes and different groups of SAR data from different angles of view. The experimental analysis and analysis of the classical SIFT method, SIFT-OCT method and the new method are carried out. In contrast, the results show that the method is superior to the SIFT method in performance, and the SIFT-OCT method can further improve the potential of performance. The sixth chapter summarizes the work of this paper, and looks forward to the next research work of the local feature extraction method of SAR images.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN957.52

【相似文献】