高分辨率遥感影像道路目标智能识别方法研究

发布时间：2018-02-26 09:43

本文关键词： 高分辨率遥感影像道路中心线提取多特征融合　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,随着多平台遥感技术与数字图像处理技术的快速发展,高分遥感技术已经成为了精确获取地表信息的重要手段之一。道路网络作为基础地理信息的重要组成部分,在车辆导航,国土规划等方面具有非常重要的作用。传统的基于高分辨率遥感影像的道路提取技术中通常需要辅以人工目视解译,该种方式虽能够精确获取道路网络信息,但是其效率不高,智能化程度较低,无法满足道路网络信息自动快速更新的需求。因此开展高性能、智能化的道路网提取方法研究具有重要的现实意义。目前,基于高分辨率遥感影像的道路提取技术一般包括道路二值图像提取,道路中心线提取两大步骤。由于高分辨率遥感影像上下文特征复杂,同谱异物,同物异谱等现象更为严重,从而造成现阶段自动化提取道路网络尚存在以下不足:(1)基于像素层次的道路提取算法,容易产生"椒盐"和"空洞"现象,基于对象层次的道路提取算法,容易产生"粘连"现象,这两种现象的存在较大程度上影响了道路网络的后续优化步骤;(2)初始提取的道路图像中,非道路区域的存在影响了道路目标的唯一性与完整性,而现有非道路区域去除算法存在鲁棒性和普适性较差的问题,需要利用较多的特征和进行复杂的参数设置;(3)现有道路中心线提取效果尚存在以下不足之处,其一从影像本身的角度出发,由于高分影像道路存在的车辆、建筑物阴影、同谱异物等情况,容易产生"断裂"的现象,其二从中心线提取算法的角度,现有道路中心线提取算法容易产生"毛刺"、拟合不准确等现象。针对现阶段道路提取存在的不足之处,本文主要完成了以下3个方面的研究内容:(1)从影像底层的特征提取角度出发,发展了顾及像素级空间特征和对象级光谱特征的多层次特征融合算法,从而有效弥补了单一层次信息的不足,从特征提取的角度来提高道路提取精度,有效改善了像素级道路提取存在的"椒盐"与对象级道路提取存在的"粘连"现象;(2)针对现有非道路区域去除算法,存在过多特征组合与参数设置,本文设计了一种新的形状描述算子,并基于OSTU自动阈值法实现了非道路区域的自动过滤,提高了道路提取算法的自动化程度;(3)针对现有道路中心线算法容易产生的"毛刺"与拟合不准确现象,引入了计算机视觉中的张量投票算法,有效地解决了上述问题,且实现了断裂道路的自动连接,最终提高了道路目标的完整程度。总的来说,本文构建了一种稳健的多特征融合方法,并实现了非道路区域的自动过滤,同时提高了道路网络形成时的完整性水平,最终形成了一套完整的高分辨率遥感影像道路提取算法。最后,本文设计了多组实验进行了算法验证,并通过与现有道路提取算法进行对比,无论从定性的角度还是定量的角度,本文算法均取得了更优的结果,论证了本文算法的有效性,同时也反映出本文算法具有更高的鲁棒性。
[Abstract]:In recent years, with the rapid development of multi platform remote sensing technology and digital image processing technology, high resolution remote sensing technology has become an important means of obtaining accurate ground information. The road network as an important part of basic geographic information, in vehicle navigation, plays a very important role in land planning and so on. High resolution remote sensing image road extraction technology usually needs to be complemented by visual interpretation based on the traditional ways, although the road network can accurately obtain information, but its efficiency is not high, low intelligence, unable to meet the road network information automatic rapid updating needs. Therefore to carry out high performance, intelligent road network extraction method research has important practical significance at present, the road extraction technology based on high resolution remote sensing images generally include road two value image extraction, extraction of the central line of the road two Step. As the context of high resolution remote sensing image features of complex, same spectrum, spectral variety phenomenon is more serious, resulting in current automatic road network extraction still exist in the following aspects: (1) extraction algorithm of pixel level based road, prone to "salt and pepper" and "empty" phenomenon, the extraction algorithm of the object hierarchy based on the road, prone to "adhesion" phenomenon, the two phenomena have greater influence on the subsequent optimization steps of road network; (2) the initial extraction of road image, non Road area has affected the uniqueness and integrity of the road, and the existing non road region removal algorithm are robust and universality of the problem of the poor, need to use more complex features and parameters; (3) the effect of the existing road centerline extraction still has the following shortcomings, one from the perspective of the image itself, Due to the presence of high image road vehicles, shadows, same spectrum etc., can easily lead to "break" phenomenon, the extraction algorithm from the angle of the center line, the existing road centerline extraction algorithm is easy to produce "burr", the fitting is not accurate and so on. Aiming at the deficiencies at this stage are the main road extraction. Research on the following 3 aspects: (1) from the point of view of the underlying image feature extraction of the development of multi level feature for pixel level spatial and spectral characteristics of the object level fusion algorithm, so as to effectively compensate for the lack of a single level of information, from the angle of feature extraction to improve the extraction accuracy of road, effective to improve the pixel level road extraction of "salt and pepper" with the object level of road extraction has "adhesion" phenomenon; (2) according to the existing non road region removal algorithm, there are too many feature combination and design parameters Set, this paper designed a new shape descriptor, which is based on the OSTU automatic threshold method to realize the automatic filtering of non Road area, improve the degree of automation of the road extraction algorithm; (3) according to the existing road center line algorithm is easy to produce "flash" fitting and inaccurate phenomenon, introduced the tensor voting algorithm of computer in the vision, to effectively solve the above problem, and automatically connecting the broken roads, and ultimately improve the integrity of the road target. In general, this paper constructs a robust fusion method of multi feature, and realizes the automatic filtering of non Road area, and improve the integrity level of road network when, eventually forming a complete set of extraction algorithm of high resolution remote sensing image. Finally, this paper designed several experiments are performed to verify the algorithm, and the algorithm and the extraction of roads Comparing with qualitative and quantitative methods, the algorithm achieves better results, and demonstrates the effectiveness of the algorithm. It also shows that our algorithm is more robust.

【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】