PSO-based procedure to find number of clusters and better initial centroids for K-means algorithm: Image segmentation as case study

Abstract

In this paper, we propose a combination of K-means algorithm and Particle Swarm Optimization (PSO) method. The K-means algorithm is utilized for data clustering. On one hand, the number of clusters (K) should be determined by expert or found by try-and-error procedure in the K-means algorithm. On the other hand, initial centroids and number of clusters (K) are influenced on the quality of resulted grouping. Therefore, the aim of the proposed procedure is using PSO and the Structural Similarity Index (SSIM) criterion as a fitness function in order to find the best value for K parameter and better initial clusters' center. Due to different value of K parameter, the number of initial centroids which should be produced is variant. Thus, length of particles in PSO method may be different in each iteration. Experimental results show the superiority of this approach in comparison with standard K-means algorithm and both of them are evaluated on image segmentation problem.