Indian Journal of Science and Technology

Abstract

Image block clustering is important in several exploratory applications such as image segmentation, image pattern classification, image compression and the like. Clustering the blocks of an image into meaningful groups to reveal useful information is a challenging problem. This paper uses Partitional algorithm for image block clustering and also clusters are arranged in a hierarchical structure to explore the different aspects of the data. K Means algorithm is a widely used Partitional algorithm. But due to its gradient descent nature, the result of this algorithm is very sensitive to the initial cluster centroids which do not produce unique clustering results every time for the same input. Many initialization methods have been proposed to address this problem, but it produces results at the cost of high computational time. A Deterministic Centroid Initialization Method (DCIM) was proposed in our earlier work for K means Clustering algorithm which was used to cluster the image blocks for content adaptive image compression. In this paper, we extend the performance analysis part of DCIM in conventional K Means and Fuzzy C Means algorithm. The performance analysis has been done with the measures such as Root Mean Square Error (RMSE), Number of iterations and CPU time. The strength of this DCIM method is that the clustering algorithms require less number of iterations to attain convergence and in producing the unique better clustering result in a single run. Clustering algorithms with DCIM was tested on a variety of images to show its strength. The experimental results show that the clustering algorithm with DCIM outperforms the clustering algorithms with Random Centroid Initialization Method (RCIM) in terms of RMSE and number of iterations. From an average performance view, K Means with DCIM produces a decrease of 20.87% in RMSE with marginal increase of 0.27 seconds at CPU time than K Means with RCIM. The validation results prove that the DCIM guarantees unique better results within a lesser number of iterations with less computational effort.
Keywords: Cluster Centroid Initialization, Fuzzy C Means, Image Block Clustering, K Means, Partitional Clustering