Publication:
Automatic number of clusters determination of clustering algorithms for image segmentation application

Loading...
Thumbnail Image
Date
2012-06-01
Authors
Tang, Jing Rui
Journal Title
Journal ISSN
Volume Title
Publisher
Research Projects
Organizational Units
Journal Issue
Abstract
Image segmentation involves a process of partitioning a natural image into regions with homogeneous texture. Due to their ease of implementation, efficiency and capability in providing promising solution in many applications, the conventional clustering algorithms, K-Means and Fuzzy C-Means clustering algorithms are widely used in image segmentation. However, these clustering algorithms are sensitive to the initialization condition of number of clusters, where the number of clusters has to be provided by the user. In this study, an enhanced version of image segmentation technique based on histogram analysis is proposed. Four histogram smoothing techniques, namely local regression technique, robust version of local regression technique, low pass filter and Savitzky-Golay smoothing filter are applied on the histogram of the input image to determine the number of clusters automatically. Both qualitative and quantitative analyses prove that histogram smoothing techniques are capable of providing the optimum number of clusters for conventional clustering algorithms. Experimental results show that for K-Means clustering algorithm, the best segmented images are produced when the number of clusters is supplied by the histogram smoothed using the robust version of local regression technique. On the other hand, the optimum number of clusters for Fuzzy C-Means clustering algorithm is produced when low pass filter is used to smooth the histogram. The resultant segmented images produced by the proposed algorithms have successfully enhanced the contrast, preserved high amount of details and produced more homogeneous regions. These findings suggest that the integration of histogram smoothing techniques with the conventional clustering algorithms to automatically find the number of clusters in the image without prior knowledge on the image has high potential for future research.
Description
Keywords
Citation