【作者】 陈爱军；

【导师】 李金宗；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2008， 博士

【摘要】 随着遥感图象分辨率的提高,图象数据量有了爆炸式的增长。如何从海量图象数据中及时、准确地获取所需信息并加以利用,成为需要解决的重大问题。如果完全靠传统的人工方法对大幅面高分辨率遥感图象进行判读和识别,效率和精度都无法保证。因此,利用计算机对遥感图象中的感兴趣目标进行自动识别,辅助判读人员完成判读工作具有非常重要的意义。本文在对现有典型人造目标自动识别技术进行分析的基础上,以感兴趣的大中型水上桥梁目标和圆形目标群为研究对象,深入研究了其在复杂背景下大幅面高分辨率遥感图象中的目标特性并实现了目标的快速筛选与识别,取得了一些有价值的研究成果。本文对遥感图象的目视判读、自动判读和交互判读进行了简要介绍并对它们的优缺点进行了分析,然后对高分辨遥感图象数据的特点进行了总结并指出开展高分辨率遥感图象目标自动识别进行辅助判读的必要性和紧迫性,最后对高分辨遥感图象中的道路、建筑物、桥梁和机场等人造目标识别技术进行了分析,提出了存在的主要问题。针对现有桥梁识别算法的不足,建立了一种基于知识的水上桥梁识别算法。根据桥梁在高分辨遥感图象中的功能特征、灰度特征、几何特征和空间上下文特征,建立了水上桥梁的先验知识库;根据高分辨率遥感图象中河流的灰度特征和形状特征利用模糊分割和聚类的方法对河流区域进行粗分割,并根据桥梁与河流之间的空间关系提出了一种确定感兴趣区域的方法,利用该方法自动提取出感兴趣区域的子图象,然后对感兴趣区域子图象进行分割、轮廓跟踪、直线拟合等操作获取潜在水上桥梁的特征,最后再根据水上桥梁的先验知识对潜在桥梁进行验证。在水上桥梁识别的过程中,河流和桥梁的先验知识始终贯穿其中,从而大大提高了检测和识别速度。通过对高分辨遥感图象中的圆形目标特征进行详细分析,建立了一种基于圆检测技术和空间分布关系的圆形目标群识别方法。该方法充分利用圆形目标群中的单一目标在外观上表现为圆形以及圆形目标集群分布的特点,利用圆检测技术从图象中检测出候选圆目标,然后利用候选圆目标的空间分布和数量关系来确认各个圆形目标群区域,在消除虚警的同时,完成了圆形目标群的识别。为了从图象中快速提取出圆形目标,本文在RCD(Randomized Circles Detection)算法的基础上,对采样点的选取方法以及证据积累过程中与像素有关的运算方法进行改进,提出了一种改进的随机圆检测算法;另外,根据圆形目标的几何特征,提出了两种快速圆检测算法;并对上述三种圆检测算法进行了对比实验。实验分析表明,本文建立的自动识别复杂背景下大幅面高分辨率遥感图象中水上桥梁和圆形目标群这两类典型人造目标的方法是快速而有效的,可辅助判读人员快速完成人造目标的判读筛选工作,具有重要的现实意义。更多还原

【Abstract】 At present, the image data are dramatically increased with the resolution of remote sensing imagery being higher. It has become an important issue how to timely acquire the useful information from abundant image data. The efficiency and precision of interpretation and recognition from large breadth remote sensing images with high resolution are not assured if the traditional manual methods are only applied. It has important value for assisting interpreters to accomplish imagery interpretation by recognizing automatically targets in remote sensing imagery with computers. In this dissertation, the recognizing and locating techniques for man-made targets such as bridges and cirlular target clusters are investigated based on existed ATR techniques. The target features and recognition methods of typical man-made objects in high resolution satellite imagery with large size and complicated background is deeply investigated, and some valuable results are achieved.The advantages and disadvantages of visual interpretation, automatic interpretation and interactive interpretation are discussed. The characteristics of data from high resolution remote sensing imagery are summarized and the necessity of developing aided interpretation is pointed out. The main shortcomings of recognition techniques of man-made objects in high resolution remote sensing imagery are presented.A knowledge-based recognition algorithm of bridge over rivers is presented in order to overcome the shortcomings of existed algorithms. Firstly, the repository of bridges over rivers is set up in terms of the function features, gray features, geometric features and spatial context features. Secondly, river areas are coarsely segmented based on fuzzy theory and clustering analysis. Moreover, a method of selecting regions of interest (ROI) is proposed in terms of the spatial relationship between rivers and bridges. With the method, sub-images are automatically obtained. Lastly, image segmentation, contour tracking and line fitting are implemented to extract candidate bridges in each sub-image and every candidate bridge is verified using priori knowledges of bridges. During the process of bridge recognition, priori knowledges have been used all the time to speed up detecting and recognizing. By analyzing the features of cirlular-target clusters in high resolution satellite remote sensing imagery, the recognition method of cirlular target clusters is presented based on circle detection techniques and spatial distribution relationship of circular objects. With the characteristics that circular objects always locate together in each target cluster, possible circular objects are firstly detected using circle detection technologies. Then, the regions of circular-target clusters are validated using the relationships of spatial distribution and quantity of possible circular objects. Cirlular target clusters are recognized by eliminating false circular objects. In order to rapidly extract circular-target culsters, an improved circle detection approach is proposed based on RCD (Randomized Circles Detection) algorithm, which has ameliorations in two ways of sampling pixel selecting and operation on pixels in evidence-collecting process. In addition, two new algorithms are presented to quickly detect circles. In the end, three algorithms are compared with one another by experiments.The experimental analyses show that the methods proposed in this dissertation for automatically recognizing bridges and circular-target culsters in high resolution satellite imagery with large breadth and complicated backgrounds are fast and effective. It has the important realism sense for interpreters to quickly interpret and screen out man-made objects.更多还原