基于多目标聚类和选择集成的SAR图像变化检测方法

发布时间：2018-03-14 23:06

本文选题：变化检测　切入点：合成孔径雷达(SAR)　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：遥感图像变化检测是通过对不同时间获得的覆盖同一区域的两幅或多幅遥感图像进行观测分析,对比得到图像之间的差异,进而检测出该地区的地物随时间发生的变化信息。目前图像变化检测算法的研究方法一般流程是,首先生成两时相遥感图像的差异图像,然后对差异图像进行分析处理,将差异图像分成变化类和未变化类两类,得到最终的检测结果。聚类就是其中对差异图进行分类的方法之一。本文针对已有聚类技术在图像分割中的不足,提出了一种无监督的基于于非支配邻域免疫算法的多目标模糊聚类算法和选择性集成策略的SAR图像变化检测算法:1.首先,提出了一种多目标聚类的差异图分析算法。算法设计两个互补的聚类目标函数评价聚类性能,目标二中引入邻域像素与中心点像素之间的灰度差和欧式空间距离加权作为聚类算法的相似性度量,在聚类过程中结合空间邻域信息。算法相比于传统的单目标聚类算法,可以更好地去除斑点噪声对聚类结果的影响,又不会造成细节的丢失。同时由于建立两个目标,避免了参数选择困难的问题。实现了在图像分割或分类过程中既保持细节的完整增强了聚类性能,同时抑制斑点噪声的目标。使用的进化多目标方法用随机产生的初始抗体种群代替初始的聚类中心,降低了传统聚类分割方法对初始聚类中心的敏感度,使用单一目标进行聚类运行多次才能生成不同的解,利用多目标优化方法运行一代即可得到。2.本文引入选择性集成策略,将多目标聚类的结果视为不同权值的同态分类器分类的结果,对初步分类的结果进行选择性集成,得到比单个聚类更好的结果。本文所提出的选择性集成策略,首先将所有聚类结果进行简单的投票集成;然后根据一次集成的结果作为评判标准,对各基分类器进行排序;排序后,选择前10%~30%进行集成,最终获得一组整合的分割结果。使用多目标优化聚类产生的结果是一组非支配解集,得到的是一组聚类中心,由聚类中心可以得到不同的分割结果。从多目标优化角度来讲,结果之间相互支配,没有优劣性可言。实际应用时,可以根据实际需要或偏好来选择其中一个解。其他大多数的多目标聚类算法是采用第三方选解策略来进行选解,其实质相当增加聚类目标。另外一些选解策略则需假设10%的实验数据已知,这对于图像聚类分割并不实际。而使用选择性集成策略进行处理,无需先验知识或第三方指标。最后的实验结果表明,相比现有的其他办法,本文所采用的方法整体取得较好的变化检测差异图分割结果。
[Abstract]:Remote sensing image change detection is obtained through observation and Analysis on different time covering the same region of the two or more images of remote sensing images, comparing the difference between the images, and then detect the change information of the region features over time. The current image change detection method is the general process is to first generate two when the phase difference between remote sensing image, and then the difference of image analysis and processing, the difference image is divided into change and no change two classes, to get the final detection results. The difference of graph clustering is one way of classifying. Aiming at the shortcomings of the existing clustering techniques in image segmentation, proposed a SAR the change of image based on unsupervised multi target nondominated neighbor immune algorithm of fuzzy clustering algorithm and selective ensemble strategy detection algorithm: 1. first, this paper presents a multi-objective Algorithm analysis of the difference graph clustering algorithm design two clustering objective function evaluation of the complementary clustering performance, the second goal is introduced between the gray and neighborhood pixels in the center pixel difference and the Euclidean distance as the similarity weighted clustering algorithm with spatial neighborhood information measure, in the process of clustering. The single objective algorithm compared with the traditional clustering algorithm, can better remove the influence of speckle noise on the clustering results, and will not cause the loss of details. At the same time due to the establishment of two goals, to avoid the difficulty of parameter selection problem. The image segmentation and classification process can keep the full details to improve the clustering performance, while suppressing speckle noise. The evolution of multi target the target used by the initial cluster center instead of the initial antibody population randomly generated, reducing the traditional clustering segmentation method of initial clustering center sensitivity Sensitivity, using a single target clustering operation several times to generate different solutions, using multi-objective optimization method to run a generation of.2. can be obtained by selective integration strategy, will result in multi objective clustering results as the classification of different weights of the homomorphic classifiers, the preliminary classification results of selective integration than single clustering better results. Selective integration strategy proposed in this paper, firstly all clustering results are simple voting integration; then according to the results of integrated as a judgment standard, for each base classifier in order; after sorting, selection of 10%~30% before integration, eventually obtain a set of integration results using multi target segmentation results. The optimal clustering is a set of non dominated solutions, get a set of cluster centers, get different segmentation results by the clustering center. From the angle of multi-objective optimization The degree of speaking, mutual control between results, no merits at all. In practical application, can choose one of the solutions according to the actual needs or preferences. Most of the other multi-objective clustering algorithm is used to select third party solution strategies to select the solution, in fact a considerable increase in poly target class. In addition some experimental data in solution selection strategy you need to assume that the 10% known, this is not practical. For image clustering segmentation using selective ensemble method for processing, without prior knowledge or third indicators. The experimental results show that, compared with other existing methods, this method has better overall results to detect the changes of image segmentation.

【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

【参考文献】