一种基于离群点信息的新型无监督聚类方法

在图像检索领域，聚类分析技术有着广泛应用。因为在对图像进行聚类分析时，通常缺少可资利用的先验知识，所以需要采用无监督的聚类算法。为了适应图像检索的需要，提出了一种新型的无监督聚类方法，即根据离群点信息来自动确定聚类算法的终止时机。此方法还弥补了现有聚类算法在离群点识别、使用上的缺欠。为验证其可行性，用其改进了CURE和ROCK两个经典算法。实验表明，改进后的两个算法都能自动终止，并能取得优于以往的聚类效果。

A New Unsupervised Clustering Method Based on Ontlier Information

There are various applications of clustering analysis techniques in the field of image retrieval. For the lack of valuable prior knowledge in the image retrieval process, unsupervised clustering algorithms should be applied. This paper proposes a new unsupervised clustering method: clustering algorithms will automatically stop according to the outlier information. This method also complements the shortages of current clustering algorithms in outlier detection and using. To show its feasibility, the paper proposes several improvements on two classical clustering algorithms, CURE and ROCK. The empirical results show that by using new method, these two algorithms can stop automatically and also achieve better performance.