Abstract: The traditional Kernel Possibilictic C-Means（KPCM） only consider the relationships within the class without enough attention to the distance between classes. When it comes to fuzzy boundary data, misclassification problems in boundary may easily occur due to the overlapping of the centers. To solve the above problems, this paper introduces a maximum penalty term between classes in high-dimensional feature space and the control parameter[λ]based on the KPCM. The new proposed algorithm which constructs a new objective function is called the Maximum center interval Kernel Possibilistic C-Means（MKPCM） clustering algorithm. The algorithm makes the distance between the centers maximum by the maximum penalty term between centers and through the control parameter[λ], it effectively avoids the event of too close centers or even overlaps. Numerical experimental results demonstrate its favorable performance especially in the matter with fuzzy boundary. In addition, it shows distinct advantages in the application of image segmentation compared to the traditional cluster methods.

Key words: Kernel Possibilistic C-Means（KPCM）, fuzzy boundary, maximum penalty term between centers

摘要： 传统核可能性C均值（KPCM）算法仅考虑类内的紧密性而忽略了类间的距离关系，在对边界模糊的数据进行聚类分析时，会引起因聚类中心距离小或重合引起的边界点误分问题。为解决上述问题，在核可能性C均值基础上引入高维特征空间中的类间极大惩罚项和调控因子[λ]，构造了全新的目标函数，称为极大中心间隔的核可能性C均值（MKPCM）聚类算法。该算法通过类间极大惩罚项使类间距离极大化，并利用调控因子[λ]合理控制类间距，较好地避免了类中心间距离小或重合的现象。通过大量的实验证明，算法对于边界模糊的数据聚类效果优于传统的聚类算法;在图像分割的实际应用中，算法也明显优于传统的聚类算法。

关键词: 核可能性C均值, 边界模糊, 类间极大惩罚项