摘要
1,引言
图像分割的目的是将图像划分为一些互不重叠的区域,这是计算机视觉中的一个重要研究领域,也是图像理解的基础.在众多的图像分割技术中[1],特征空间聚类可以说是最常用的方法之一.通常用一确定的特征表征同一分割区域的像素.这些特征被量化成特征变量,同一分割区域的像素的特征变量基本上有类似的数值,不同分割区域的像素特征变量数值不同.在实施图像分割时,首先在特征空间把特征变量聚类,然后把特征空间的每一点映射回到图像空间的像素.
When image segmentation is treated as a problem of clustering pixels, statistical finite mixture model can be used to classify image samples. With estimated mixture model parameters, we can use some Information Theoretical Criteria to determine how many regions should be segmented on a given image without a priori knowledge to conduct automatic image segmentation. In this paper, we consider the problem in practical implementing segmentation based on mixture models and suggest combining several techniques such as data reduction. Competitive Learning and variant EM algorithm to effectively estimate mixture parameters and reduce intensive computation task. The combined technique is of significance for automatic image segmentation in real-time or near real-time applications.
出处
《计算机科学》
CSCD
北大核心
2002年第8期101-103,共3页
Computer Science
