基于克隆选择和聚类的遥感图像分割研究

发布时间：2018-02-23 06:02

本文关键词： 高分辨率遥感图像分割克隆选择模糊C均值聚类谱嵌入式聚类核函数　出处：《中国矿业大学》2014年博士论文　论文类型：学位论文

【摘要】：随着遥感技术的不断发展，利用新的遥感平台获取的高分辨率图像不但光谱信息丰富，同时包含着大量地物表面更多的形状、纹理等细节信息。但是随之而来的挑战是如何对这些数据进行有效的处理，为后续的具体应用提供支持。遥感图像分割是数据分析理解前的关键步骤，图像分割效果的好坏将直接影响到后续的目标特征提取描述、识别与分类。该文围绕提高高分辨遥感图像的分割性能这一中心环节，在对图像分割研究中使用的克隆选择算法（Clonal Selection Algorithm，CS）、谱嵌入式聚类方法（Spectral Embedded Clustering，，SEC）以及模糊C均值聚类（Fuzzy C-meansClustering，FCM）进行分析比较的基础上，结合遥感图像的特点从理论、方法上进行改进，提出了相应的改进算法，利用QuickBird高分辨率遥感图像对相关分割算法进行了试验、评价和对比。该文主要研究工作如下：首先，通过深入分析人工免疫理论中的克隆选择机理，针对基本克隆选择算法（Clonal Selection Algorithm）存在的不足，通过增加交叉操作及根据抗体浓度调节种群规模，提出了改进的克隆选择算法，达到了提高抗体多样性，提高算法全局搜索能力的目标。通过与二维最大熵和多重空间构造图像分割方法相结合，利用实验证明了改进算法优于基本克隆算法。其次，针对谱聚类算法（Spectral Clustering，SC）和谱嵌入式聚类的缺点，引入核函数设计了基于核函数的谱嵌入式聚类算法（Kernel Function-based SpectralEmbedded Clustering，KSEC），使用三类核函数对算法进行了构造和实现，将谱嵌入式聚类和基于核函数的谱嵌入式聚类算法应用于遥感图像的分割，提高了分割的精度。第三，在对模糊C均值聚类（Fuzzy C-means Clustering，FCM）及改进模糊C均值聚类算法进行研究的基础上，引入图像相邻像素之间的空间引力概念，适度纳入局部空间信息和灰度信息，提出了一种基于空间引力的模糊局部信息C均值聚类（Neighborhood-Attraction-Based Fuzzy Local Information C-means Clustering，FLNAICM），克服了图像中相邻像素对中心像素的影响及噪声对分割结果的影响问题，成功应用到遥感图像分割中，并取得了比前2种算法更好的分割效果。最后，通过对3种改进算法对遥感图像分割精度的对比，证明对具有模糊特性的遥感图像进行分割时，基于模糊理论的聚类分割方法更为有效。
[Abstract]:With the development of remote sensing technology, the high-resolution images obtained by the new remote sensing platform are not only rich in spectral information, but also contain more shapes on the surface of a large number of ground objects. However, the challenge is how to deal with these data effectively and provide support for subsequent applications. Remote sensing image segmentation is a key step before data analysis and understanding. The effect of image segmentation will directly affect the target feature extraction description, recognition and classification. This paper focuses on improving the segmentation performance of high resolution remote sensing images. Based on the analysis and comparison of Clonal Selection algorithm, Spectral Embedded clustering algorithm and Fuzzy C-Means clustering algorithm used in image segmentation, this paper improves the theory and method of remote sensing image combining with the characteristics of remote sensing image, the spectral embedded clustering method (Spectral Embedded clustering algorithm) and fuzzy C-means clustering algorithm (FCM). An improved algorithm is proposed, and the correlation segmentation algorithm is tested, evaluated and compared with QuickBird high resolution remote sensing image. The main work of this paper is as follows:. First of all, by analyzing the mechanism of clone selection in artificial immune theory, aiming at the shortcomings of basic clone selection algorithm, by increasing cross-operation and adjusting population size according to antibody concentration, An improved clonal selection algorithm is proposed to improve the diversity of antibodies and improve the global search ability of the algorithm. The experimental results show that the improved algorithm is superior to the basic clone algorithm. Secondly, aiming at the shortcomings of spectral clustering algorithm (Spectral clustering) and spectral embedded clustering, a kernel function based spectral embedded clustering algorithm, Kernel Function-based SpectralEmbedded clustering algorithm, is designed, and the algorithm is constructed and implemented using three kernel functions. The spectral embedded clustering algorithm and the spectral embedded clustering algorithm based on kernel function are applied to the remote sensing image segmentation, which improves the segmentation accuracy. Thirdly, based on the research of fuzzy C-means clustering and improved fuzzy C-means clustering algorithm, the concept of spatial gravity between adjacent pixels is introduced, and local spatial information and gray level information are appropriately incorporated. In this paper, a fuzzy local information clustering method based on spatial gravity is proposed, which is based on fuzzy local information, Attraction-Based Fuzzy Local Information C-means clustering algorithm. It overcomes the influence of adjacent pixels on center pixels and noise on segmentation results, and is successfully applied to remote sensing image segmentation. The segmentation effect is better than the first two algorithms. Finally, by comparing the accuracy of remote sensing image segmentation with three improved algorithms, it is proved that the clustering segmentation method based on fuzzy theory is more effective when the remote sensing image with fuzzy characteristics is segmented.
【学位授予单位】：中国矿业大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP391.41

【参考文献】