PARALLEL K-MEANS CLUSTERING WITH NAÏVE SHARDING FOR UNSUPERVISED IMAGE SEGMENTATION VIA MPI

Abstract

In digital image processing, image segmentation is an essential step in which an image is partitioned into groups of pixels. k-means clustering algorithm, which is often considered as fast and efficient, is one of the most widely used clustering algorithms to segment an image. However, as the problem size gets larger, the k-means starts to spend a significant amount of time to process. At this point, parallelization techniques should be applied to reduce the required time. Designing an efficient parallel and distributed model is not a trivial job since it should correspond to the parallel computer architecture and take communication and load balancing among processors into account. In this study, we propose a parallel and distributed k-means clustering algorithm with naive sharding centroid initialization for image segmentation. The proposed algorithm adopts the Message Passing Interface (MPI) standard to take advantage of the computational power of distributed computing nodes in a High Performance Computing Cluster. We demonstrate the parallel scalability of the proposed algorithm using up to 128 cores that achieves approximately 104 times faster clustering time.