【作者】 管慧娟；

【导师】 冯林；

【作者基本信息】 大连理工大学 ， 计算机应用技术， 2006， 硕士

【摘要】 图像分割是计算机视觉中的一个经典难题。本文对基于区域的图像分割技术进行了深入研究。在方法学上,重点研究了三类算法,即基于分水岭变换的图像分割算法、聚类算法、基于图论的图像分割算法,并提出了基于分水岭变换和核聚类的图像分割方法及融合分水岭变换和图论的图像分割方法。 分水岭算法假设二维图像是一个二维地形图,图像的灰度表示地形的高度,分水岭变换就是模拟该地形图侵入水的过程。分水岭算法有着较好的鲁棒性,但是却容易形成过分割。聚类分割算法也一直是众多图像分割方法中较常见的一种。由于图像的大数据量性质,传统的聚类分割不只相当耗时,而且有效性也有待加强。 在基于分水岭变换和核聚类的图像分割方法中,针对分水岭变换过分割的缺点,本文提出用核聚类算法进行区域的合并。核聚类的优势在于利用Mercer核函数把样本映射到高维特征空间,增加了对样本特征的优化,在高维特征空间中进行聚类,在性能上比经典的聚类算法有较大的改进。 基于图论的图象分割方法是近年来国际上图象分割领域的一个新的研究热点,但把图像中的每个像素或者特征点看作图中的一个节点,随着图中节点数的增大,问题的求解将变得异常费时。在融合分水岭变换和图论的图像分割方法中,本文把分水岭变换后的小区域看作建立的无向带权图的节点,再结合各小区域的灰度和空间信息从全局角度用Normalized Cut方法在区域之间进行分割,产生最终的分割结果,并且该方法是一种全自动的分割方法。更多还原

【Abstract】 Image segmentation is a key problem in computer vision. The region-based image segmentation method is studied in this paper. In methodology, three aspects are made research: watershed translation based image segmentation method, kernel clustering and graph theory based image segmentation technique. Two kinds of segmentation algorithm are proposed: image segmentation based on watershed transform, kernel clustering and image segmentation approach by combining watershed translation and graph theory.An image is regarded as a topographic surface. Gray values denote altitude. Watershed translation is the process that the surface is immersed in the water. Watershed algorithm provides the advantages of stabilization and speediness, but is prone to over-segmentation. Clustering algorithm is well-known segmentation method in image project. Considering large quantity data of image, classical clustering algorithms are poor in speed and accurateness.Kernel clustering algorithm is used to merging small partition separated by watershed translation in the image segmentation method based on watershed transform and kernel clustering. Kernel clustering maps the data in the original space to a high-dimensional feature space by using mercer kernel function, and optimizes the initial data. There are great improvements in performance compared with classical clustering algorithms.There has been an increasing interest in graph-theoretic segmentation algorithms based on clustering recently. An image is represented as a similarity edge-weighted graph, where the vertices represent individual pixels. The method is rather time-consuming with an increase in pixels. Small partitions are considered as vertices of an undirected weighted graph. Integrating gray feature and spatial location of each partition, normalized cut is used to segment between partitions from global view, and then produces the final segmented image. It is an effective image segmentation approach.更多还原