New Three Methods for Improving Initialization of k-Means Clustering

Abstract

The traditional k-means algorithm is a classical clustering method which widely used in variant application such as image processing, computer vision, pattern recognition and machine learning. It is known that, the final result depends on the initial starting points. Generally, initial cluster centers are selected randomly, so the algorithm could not lead to the unique result. In this paper, we present a new algorithm which includes three methods to compute initial centers for k-means clustering. First one is called geometric method which depends on equal areas of distribution. The second is called block method which segments the image into uniform areas. The last method called hybrid which combined between first and second methods. The experimental results appeared quite satisfactory.