Improvement of K-Means Clustering Through Center's Hybrid Initialization

This paper evaluates the effect of initial center selection on cluster quality and the rate of convergence in K-means clustering. In this paper, a new approach is proposed that uses random initialization and spatial distribution scores. The main objective is to enhance K-means clustering by creating a simple, novel, and efficient technique shredded for data classification and other high-level image processing tasks. Our approach relies on image color segmentation through the application of a transformation to the image's RGB color space, representing every pixel by three values of different components of color, divided into blocks, and then a calculation of the density of dots in every block. This value of density mark provided information about the distribution of color values around various parts of an image. It uses a tactical approach in selecting initial centers, namely Scored Centers Initialization, or SCI. It chooses the block with the highest score as the initial center block chosen randomly. Then, it selects subsequent centers based on the probability assessed by the Euclidean distance from the nearest one picked center and the density score of the block. It continuously does this until the desired number of centers is reached. The experimental results indicate that the clustering effect of the approach proposed was better than the traditional Random Centers Initialization method, thus proving the proposed strategy was effective. Based on the consideration of point-to-center distances and the spatial distribution score, it captured better data structures to get quicker convergence and higher quality in the cluster assignment.