A New Method for the Initialization of Clustering Algorithms based on Histogram Analysis

A. Castro, C. Bveda, and B. Arcay (Spain)

Keywords

Fuzzy Clustering; Segmentation; Fuzzy C-Means; Cluster Validity.

Abstract

Most clustering algorithms require two parameters: the number of clusters and a sample of each cluster. These parameters are critical, because their value determines the convergence speed of the algorithm and the obtained result. The technique that is most often used to calculate these values is the application of validity indices that suggest the best amount of clusters for the division of the image. Our group proposes an algorithm for the determination of the above parameters on the basis of the analysis of the histogram. The algorithm starts by smoothening the histogram, it then calculates the various valleys and peaks and on that basis the area and the distance between the different modes, and finally it uses these two parameters to provide the number of clusters and a sample of each. The results of testing various medical images show that the algorithm is able to divide the image into a correct number of clusters for its analysis, and that the provided centroids make the algorithm merge rapidly towards a solution.

Important Links:



Go Back


IASTED
Rotating Call For Paper Image