Image segmentation is a crucial and primary step in image processing, and it has numerous applications in recognition and detection. Image segmentation is performed mainly using classification and clustering. Classification requires prior information and needs operator intervention in performing segmentation. Clustering is preferred as it is unsupervised and does not require prior information. However, the clustering algorithms require initial centroids in order to obtain the clusters. The wrongly chosen clusters results in local minima producing invalid segmentation regions. In this paper a novel initial centroid selection algorithm is presented which assists the clustering algorithm to result in the region close to ground truth in limited iterations.