基于组织进化的聚类算法及其在SAR图像分割中的应用

发布时间：2018-04-30 19:27

本文选题：组织进化 + DMEA　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：合成孔径雷达(SAR)图像正在获得越来越广泛的应用,SAR传感器可以穿透云层,在恶劣的天气情况下和夜间情况下也依然能够工作,而光学传感器则做不到这些。SAR图像应用中的一个重要环节就是对图像进行正确的分割,它是实现SAR图像理解的基础,像图像变化检测、目标识别等技术都离不开良好的图像分割结果。聚类分析是一项常用的挖掘数据间关系的技术,它能够根据数据之间的相似性分布把数据集划分为数个聚类,使得在每个聚类内的数据都具有相似的属性,而不同聚类间的数据则差异较大。基于聚类分析的这一功能,把聚类分析应用于图像分割中已成为一个热门的研究方向。以下是本文在这一研究方向上的主要工作:1、传统的模糊c均值(FCM)聚类算法易陷于局部最优解,对初始值敏感,且应用于图像分割时没有考虑图像的局部信息,对噪声的抑制能力较差。为了改进上诉缺点,本文提出了基于组织进化算法的FCM聚类算法(OEA-FCM),该算法利用组织进化(OEA)的全局搜索能力解决传统FCM算法易陷入局部最优解和对初始值敏感的缺陷,同时引入了图像中的空间信息,有效的抑制了噪声的干扰。2、迭代自组织算法(ISODATA)是一个经典的聚类算法,它在K-means算法的基础上引入了对聚类的分裂操作和融合操作,从而可以自动的调整聚类中心和聚类的个数。然而该算法对分裂操作的阈值和融合操作的阈值极为敏感,且该阈值参数不易设定。为了解决这一问题,本文提出了基于组织进化的ISODATA聚类算法(OEA-ISODATA),该算法利用OEA的解空间搜索能力去搜索ISODATA聚类算法参数的最优值,同时OEA-ISODATA算法引入了像素块的策略,能够利用有效的空间信息,滤除无用的空间信息,从而提高了算法对噪声干扰的鲁棒性。3、由于单个聚类有效性指标在设计时仅考虑聚类的某一或某些方面特性,没有某一聚类有效性指标能够适用于全部数据类型,因此采用单个聚类有效性指标做为评价函数的单目标进化算法在搜索解空间时就能力有限。为了解决这一问题,我们将DMEA多目标进化算法应用于聚类分析中,提出了基于DMEA多目标进化算法的聚类算法(DMECA)。DMECA在产生新个体时同时考虑种群多样性和收敛能力,此外还构建、更新一个外部档案用于存储非支配解,此外在更新档案和种群时都尽量使得非支配解在pareto曲线上均匀的分布。
[Abstract]:Synthetic Aperture Radar (SAR) images are becoming more and more widely used to penetrate clouds and still work in bad weather and at night. But the optical sensor can not do these. One of the important links in the application of Sar image is to segment the image correctly. It is the basis of realizing the understanding of SAR image, such as image change detection. Target recognition and other techniques are inseparable from good image segmentation results. Clustering analysis is a commonly used technique for mining the relationships between data. It can divide the data sets into several clusters according to the similarity distribution between the data, so that the data in each cluster has similar attributes. But the data of different clustering are quite different. Based on the function of clustering analysis, the application of clustering analysis in image segmentation has become a hot research direction. The following is the main work of this paper: 1. The traditional fuzzy c-means FCM-based clustering algorithm is easily trapped in the local optimal solution, is sensitive to the initial value, and when applied to image segmentation, it does not consider the local information of the image. The ability to suppress noise is poor. In order to improve the defects of appeal, an FCM clustering algorithm based on organizational evolution algorithm (OEA) is proposed in this paper. The algorithm uses the global search ability of organizational evolution to solve the shortcomings of traditional FCM algorithm which is prone to fall into local optimal solution and sensitive to initial value. At the same time, the spatial information in the image is introduced, which can effectively suppress the interference of noise. The iterative self-organizing algorithm (ISODATAA) is a classical clustering algorithm. It introduces the splitting operation and fusion operation of clustering on the basis of K-means algorithm. Thus, the cluster center and the number of clusters can be automatically adjusted. However, the algorithm is very sensitive to the threshold of split operation and fusion operation, and the threshold parameter is difficult to set. In order to solve this problem, an OEA-ISODATAA clustering algorithm based on organizational evolution is proposed in this paper. The algorithm uses the spatial search ability of OEA to search the optimal value of the parameters of the ISODATA clustering algorithm, and the OEA-ISODATA algorithm introduces the strategy of pixel block. It can use the effective spatial information to filter out the useless spatial information, thus improving the robustness of the algorithm to noise interference. Because the single clustering validity index only considers one or some aspects of the clustering characteristics in the design, No single clustering validity index can be applied to all data types, so the single-objective evolutionary algorithm using a single clustering validity index as an evaluation function has limited ability to search solution space. In order to solve this problem, we apply DMEA multi-objective evolutionary algorithm to cluster analysis, and propose a clustering algorithm based on DMEA multi-objective evolutionary algorithm, which takes into account both population diversity and convergence ability when generating new individuals. An external file is updated to store the non-dominated solution, and when updating the file and the population, the non-dominated solution is uniformly distributed on the pareto curve.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】