【作者】 周家香；

【导师】 朱建军；

【作者基本信息】 中南大学 ， 大地测量学与测量工程， 2012， 博士

【摘要】 随着遥感技术的发展,对地观测系统已获取了海量各种类型遥感数据。现有研究表明,数据转化为信息存在许多不足,而遥感图像分割又是实现数据转化为信息过程中起着重要作用的一项关键技术,也是遥感图像处理领域的重点及难点课题。虽然已发展了大量的图像分割方法,并取得了一些研究成果,但应用到遥感图像分割中仍存在算法的适用性差、分割效率低、分割精度不高等不足。针对这些不足,本文拟采用Mean Shift算法进行遥感图像分割,充分利用图像的多维特征,自适应降噪的同时有效保留目标物体的边缘信息,以提高图像分割的精度和可靠性。本文研究工作主要包括：(1)提出一种结合纹理特征的自适应带宽遥感图像分割方法,以解决Mean Shift算法分割精度不高的问题。传统Mean Shift算法只使用了图像的“位置-颜色”域特征,导致图像分割精度不高。该方法则充分利用遥感图像的“位置-颜色-纹理”域组成多维特征空间,发展自适应带宽策略。先对“位置-颜色”域特征数据进行聚类,再对聚类结果计算每一聚类区域的空间带宽、灰度带宽和纹理带宽,最后在“位置-颜色-纹理”域进行自适应聚类,得到分割结果。实验结果表明,该方法具有自适应性和稳健性,可以较好地提高遥感图像分割精度。(2)引入一种区域合并方法,用于图像分割的后续处理中,以解决Mean Shift分割算法易产生的过分割问题。该方法先根据待处理遥感图像的空间分辨率确定固定空间带宽,再利用plug-in规则计算遥感图像每个波段的颜色带宽。采用基于区域面积加权的区域相似度准则和基于区域熵的合并停止准则来进行分割后的区域合并,从而解决过分割问题。(3)提出了一种改进的快速遥感图像分割方法,以解决Mean Shift算法迭代时间长不适于处理海量的遥感图像的问题。针对影响Mean Shift算法的时间复杂度的各个变量分别提出加速策略；先利用固定带宽的高斯Mean Shift算法进行聚类得到超像素,再计算每个超像素的自适应带宽；最后采用基于区域的超像素融合来完成遥感图像分割,以达到快速地分割遥感图像。(4)采用成熟的分割评价方法,用以评估遥感图像分割效果。对比分析了马丁误差估量法和对象级一致性误差估量法。实验比较得出：对象级一致性误差估量法是基于对象水平,能够很好地量化分割图像与参考图像之间的差异；相对于马丁误差估量法而言,对象级一致性误差估量法能够正确地反映分割图像中存在的过分割和欠分割状况,其评价结果与主观评价更加相符。(5)发展一种基于Mean Shift遥感图像分割的道路提取方法。该方法首先应用Mean Shift算法实现道路图像的初步分割,再合并灰度相似的区域,并依据直方图选取最佳的阈值,进行二值化分割；然后引入形状因子去除混杂在图像中与道路形状特征不相似的区域,对于仍然与道路相连的非道路区域,则构造多方向形态学滤波对其进行剔除,从而分割出独立的道路区域,同时提取出道路线；最后连接断裂的道路线,实现道路网的提取。多组实验结果表明,该方法能很好地从复杂环境中提取道路网,特别是对直线型道路尤其有效。最后,总结了本文的研究成果。下一步需要深入的研究工作有：(1)考虑分割的多尺度性,实现基于Mean Shift算法的多尺度遥感图像分割；(2)考虑利用Gabor滤波器来提取纹理特征,或将更多的特征如形状等特征用于Mean Shift遥感图像分割中。更多还原

【Abstract】 With the development of remote sensing technology, abundant and various styles remote sensing data has been obtained from current earth observation system. Existing investigations show that there are many deficiencies in the process of data being translated to information, and remote sensing image segmentation is a key technology and difficult task of remote sensing image processing field. Large numbers of segmentation approaches have been developed, and some research progeny have been gained, but when these approaches are used to remote sensing image segmentation, there are also some defects such as limited adaptability, low segmentation efficiency, and low segmentation precision. Aiming at these defects and in order to improve segmentation precision and reliability, this paper improved classical Mean Shift algorithm to segment remote sensing images by fully useing multidimension feature of remote sensing images, it can be robustness to noise adaptively, as well as preserving edge information of object target effectively. The contents of this paper mainly include:(1) In order to solve the low segmentation precision of Mean Shift algorithm in the process of segmenting remote sensing images, a segmentation approach utilizing texture features and adaptive bandwidths is proposed, Classical Mean Shift algorithm only uses the spatial-range features, and can easily lead to lower segmentation precision. While the proposed method uses the features of spatial-range-texture to form multi-dimension features spaces, and develop adaptive bandwidth strategy. Firstly data clustering is carried out in the space of position-range; then spatial bandwidth, range bandwidth and texture bandwidth of each region are calculated according to previous clustering results; lastly segmentation results are gotten by adaptive clustering in the space of position-range-texture. Experiment results show that the proposed method can improve segmentation precision of remote sensing images with high adaptability and robutness.(2) In order to overcome the over-segmentation of classical Mean Shift algorithm, a region combination method is developed to postprocess the initial over-segmentation images. Firstly, spatial bandwidth is selected according to the resolution of remote sensing images under study; then spectrum bandwidths of each band are estimated by using plug-in rules; lastly, segmented regions were merged by using regions areas weighed similarity rule and region entropy based region merge stopping rules to solve over-segmentation problem of classical Mean Shift algorithm.(3) An improved fast segmentation method is proposed in this paper, in order to solve long iterative time of classical Mean Shift algorithm, which is not apapt to mass remote sensing imageries. Aiming at fast segment remote sensing imagry, some accelerating strategies were proposed to solve each issue which affects time complication of classical Mean Shift algorithm. Firstly, super-pixels are gotten by using fixed bandwidths Gauss Mean Shift cluster algorithm. Then dandwidth of each super-pixel is calculated adaptively. Finally, remote sensing images segmentation is performed by using region-based super-pixels fusing process, and then high precision segmented results can be obtained.(4) Mature segmentation evaluation method is adopted to evaluate remote sensing image segmentation. Martin error measure method and object-level consistency error meature method are contrastively analyzed firstly. The comparison experiments show that the object-level consistency error meature method works at the object level and can effectively measure the discrepancy between a segmented image and the reference image. Compared with Martin error measure method, object-level consistency error meature method can correctly reflect over-segmentation and under-segmentation of segmented images, and its evaluation result can be consistent with the subjective evaluation much better.(5) A roads extraction method based on Mean Shift segmented remote sensing image is proposed. Firstly, road images are initially segmented by using Mean Shift algorithm, and regions with similar gray values are merged, and binarization segmentation is completed by selecting the optimal thresholds based on histogram of segmented images. Then shape indices are used to remove those regions mixed in image which have different shapes comparing to road; in order to ensure the independence of each road target candidate, a multidirectional morphological filtering algorithm is designed to separate road from the neighboring non-road objects, and then road lines are extracted. Finally, road network is extracted by connecting the broken road lines. Several experimental results show that the proposed method can be used to extract roads network from remote sensing images even under complex conditions, especially for the straight roads.At last, after concluding all research work in this paper, further work need to be in-depth studied:(1) Consider multi-scale factors of remote sensing, and realize multi-scale remote sensing image segmentation based on Mean Shift algorithm.(2) Consider extracting textures features by using Gabor filter, or use more features such as shape features to segment remote sensing images based on Mean Shift algorithm.更多还原