Abstract: Kmeans algorithm belongs to clustering methods. It is usually applied to image segmentation. To determine the best value of K, it takes Mahalanobis distance as measurement in HSI color space of image. From the perspective of information theory, a target function related with maximum weighted entropy is proposed, which converts the problem of seeking best K into optimizing the target function, thus, it can realize the unsupervised segmentation for color image. Its principle is simple and easy to implement, which can also get better results than the traditional methods.

Key words: Kmeans, HSI color space, Mahalanobis distance, maximum weighted entropy

摘要： K均值算法属于聚类方法的一种，常用于图像分割。针对如何确定最优聚类数K这一关键问题，在彩色图像的HSI颜色空间中，以马氏距离为距离测度进行K均值聚类，从信息论的角度出发，利用最大加权熵定义了一个目标函数，将最优聚类个数K的求取转换为目标函数的寻优，实现了彩色图像的无监督分割。该方法原理简单，易于实现，能获得比传统方法更好的分割效果。

关键词: K均值, HSI颜色空间, 马氏距离, 最大加权熵