【作者】 黄志坚；

【导师】 李磊；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2009， 硕士

【摘要】 随着遥感技术的发展,遥感影像的分辨率越来越高,描述地表状况的信息越来越丰富,从海量的遥感数据中挖掘对人类生产和生活有用的数据变得尤为重要,而自动提取地物属性信息成为数据生产的瓶颈。从上世纪七十年代以来,国内外相继展开了遥感图像道路提取的研究,取得了一系列的成果,但是到目前为止仍然没有一种成熟可靠的自动道路提取方法。现有道路提取方法离人类的生产生活需求还比较远。进一步研究高分辨率遥感影像的道路提取具有重要的理论和实践意义,因此本文围绕道路的半自动和自动提取的问题进行了研究。本文在深入分析了道路的现实特征和遥感影像特征的基础之上,吸收和借鉴前人的工作成果,进一步完善了道路和道路网模型,为遥感影像道路提取打下了基础。由于传感器在获取图像的过程中会受到自然的、系统的多种因素的影响,不可避免地产生了一定的噪声,为了保证系统的整体性能和稳定度,在进行道路提取之前进行适当的预处理是十分必要的。本文在对比各种平滑滤波方法的基础上选定了不损害边缘信息的中值滤波。为了增强道路与背景的对比度,采用了分段线性增强。后续的实验表明这些预处理方法是行之有效的。图像配准技术相对于道路提取技术成熟得多,为了充分利用图像配准的思想和框架,本文结合模板匹配方法和图像配准框架,提出了一种新的半自动的高分辨率遥感影像道路提取方法。该方法通过种子点自动获取道路影像模板,并采用扩展的Kalman滤波方法进行预测,在预测点附近选取目标窗口,通过最速梯度下降法在目标窗口内快速搜索与模板最相似的影像,如此往复,进而实现道路追踪。实验证明该方法具有一定的越障能力,需要的人工干预少,能够较准确地提取农村和城郊道路。自动提取道路是本文的重点。传统的线段检测方法,对于弧形边缘没有充分地考虑,而弧线是道路的重要线性特征之一。本文改进了边缘检测方法,并运用感知编组的理论指导线段的合并和生长,提取出较完整的道路网。实验证明该方法能够快速、准确、自动地提取出较高等级的道路。最后,本文简要介绍了遥感影像道路生产系统的设计与实现。该系统具有开放性、兼容性好和人机交互友好等特点。更多还原

【Abstract】 With the development of remote sensing technology, the resolutionof remote sensing images becomes higher and higher, and the information for description of the surface of the globe becomes more and more abundant. So how to extract the useful data for human’s generation and life from masssive remote sensing data becomes very important. But extracting ground object attribution has become the bottleneck of data generation. Since from the 1970’s, researchers of home and abord attempt to find the way of extracting road from high resolution remote sensing image successively, and have obtained a series of achievements, but there isn’t any mature and reliable automatic method up to now. These achievements are far away from human’s generation and life. So there is important theory and practical meaning to further explore the way of extracting roads from high resolutionof remote sensing images. This paper shows the attemption around roads extraction semi-automaticly and automaticly.Based on the analysis of the real feature and image feature of roads, with assimilation and utilization of the achievements of previous, we improved the roads and roadnet model. The model is useful for the subsequent work. Because of the interference from the system of sensor and nature, noise is inevitable. Proper pretreatment before extracrion is necessory. This paper choosed the median filter for not losing endge imformation, based on comparation of some classic method. We choosed local linear enhancement to enhance contrast. The sebquention work shows this pretreatment is effective.The technology of image registration is more matuer. To use the method and framwork of image registration, we proposed a new way to extract roads semi-automatically. Getting image template by seed, and using Kalman filter to predict, we got the target window. Searching the most similar image as template in target window by using gradient descent algorithm, we get the next road point. Do the same thing as above, we realized road tracing. Experiment, has showed that this way has the ability of overleap obstacles, needs less human Interference, and can extract road from rural aera extractly.Automatic-road-extraction is the major of this paper. The traditional line segemernt detection hasn’t considered about curve enoughly, but curve is an important part of road. By improving the line segemernt detection method, using perceptual organization theory to guide line segement mergeing and growth, we can get relatively complete roadnet. Experiments have showed that this way can extract relatively high-level road quickly, extractly and automatically.At last, this paper introduced the designation and implementation of the Road Data Generation System of Remote Sensing Image (RDGSRS). This system has the character of good compatibility, friendly human-computer interaction and open platform.更多还原