An Efficient Selection of Initial Centroids for K-Means Clustering

Abstract

Abstract: One of the most popular unsupervised clustering algorithms is the K-Means clustering algorithm which can be used for segmentation to analyse the data. It is a centroid-based algorithm, where it calculates the distances to assign a point to a cluster. Each cluster is associated with a centroid. The selection of initial centroids and the number of clusters play a major role to decide the performance of the algorithm. In this context, many researchers worked on, but they may not reach a goal to cluster the images in minimum runtime. Existing histogram based initial centroid selection methods are used on grayscale images only.  Two methods, i.e., Histogram based initial centroids selection and Equalized Histogram based initial centroids selection to cluster colour images have been proposed in this paper.

The colour image has been divided into R, G, B, three channels and calculated histogram to select initial centroids for clustering algorithm. This method validated on three benchmark images and compared to the existing K-Means algorithm and K-Means++ algorithms. The proposed methods give an efficient result compared to the existing algorithms in terms of runtime.

Index Terms: Histogram, Equalized Histogram, Initial Centroids, K-clusters, K-Means clustering, K-Means++ clustering.