Optimized K-means (OKM) clustering algorithm for image segmentation

This paper presents the optimized K-means (OKM) algorithm that can homogenously segment an image into regions of interest with the capability of avoiding the dead centre and trapped centre at local minima phenomena. Despite the fact that the previous improvements of the conventional K-means (KM) algorithm could significantly reduce or avoid the former problem, the latter problem could only be avoided by those algorithms, if an appropriate initial value is assigned to all clusters. In this study the modification on the hard membership concept as employed by the conventional KM algorithm is considered. As the process of a pixel is assigned to its associate cluster, if the pixel has equal distance to two or more adjacent cluster centres, the pixel will be assigned to the cluster with null (e. g., no members) or to the cluster with a lower fitness value. The qualitative and quantitative analyses have been performed to investigate the robustness of the proposed algorithm. It is concluded that from the experimental results, the new approach is effective to avoid dead centre and trapped centre at local minima which leads to producing better and more homogenous segmented images.