S. N. Sulaiman and N. A. M. Isa: Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation 26 Adaptive Fuzzy-K-means Clustering Algorithm for Image Segmentation Siti Noraini Sulaiman and Nor Ashidi Mat Isa, Member, IEEE Abstract — Clustering algorithms have successfully been applied as a digital image segmentation technique in various fields and applications. However, those clustering algorithms are only applicable for specific images such as medical images, microscopic images etc. In this paper, we present a new clustering algorithm called Adaptive Fuzzy-K-means (AFKM) clustering for image segmentation which could be applied on general images and/or specific images (i.e., medical and microscopic images), captured using different consumer electronic products namely, for example, the common digital cameras and CCD cameras. The algorithm employs the concepts of fuzziness and belongingness to provide a better and more adaptive clustering process as compared to several conventional clustering algorithms. Both qualitative and quantitative analyses favour the proposed AFKM algorithm in terms of providing a better segmentation performance for various types of images and various number of segmented regions. Based on the results obtained, the proposed algorithm gives better visual quality as compared to several other clustering methods. 1 . Index Terms — Adaptive Fuzzy-K-means Clustering (AFKM), clustering, image segmentation, digital image processing. I. INTRODUCTION Clustering is a process of grouping a set of objects into classes of similar characteristics. It has been extensively used in many areas, including in the statistics [1], [2], machine learning [3]-[5], pattern recognition [6]-[8], data mining [9]- [14], and image processing [15], [16]. In digital image processing, segmentation is essential for image description and classification. The technique is commonly used by many consumer electronic products (i.e., conventional digital image) or in a specific application field such as the medical digital image. The algorithms are normally based on similarity and particularity, which can be divided into different categories; thresholding [17], template 1 This work was partially supported by Majlis Kanser Nasional (MAKNA), Kuala Lumpur, Malaysia, under the Grant “Development of an Intelligent Screening System for Cervical Cancer”. Siti Noraini Sulaiman is a PhD student at Imaging and Intelligent System Research Team (ISRT), School of Electrical & Electronics Engineering, Universiti Sains Malaysia, Engineering Campus, Penang, Malaysia (e-mail: siti.noraini.sulaiman@hotmail.com). Nor Ashidi Mat Isa is an Associate Professor, Imaging and Intelligent System Research Team (ISRT), School of Electrical & Electronics Engineering, Universiti Sains Malaysia, Engineering Campus, Penang, Malaysia (phone: +604 5996051; fax: +604 5941023; e-mail: ashidi@eng.usm.my). matching [18], [19], region growing [20], [21], edge detection [22], [23], and clustering [24]. Clustering algorithm has been applied as a digital image segmentation technique in various fields such as engineering, computer, and mathematics. Recently, the application of clustering algorithms has been further applied to the medical field, specifically in the biomedical image analysis wherein images are produced by medical imaging devices. Previous studies proved that clustering algorithms are capable in segmenting and determining certain regions of interest in medical images [25]-[32]. In a biomedical image segmentation task, clustering algorithm is often deemed suitable since the number of cluster for the structure of interest is usually known from its anatomical information [30]. Among the clustering formulations based on minimizing formal objective functions, the most widely used and studied is the K-means (KM) clustering. KM is an exclusive clustering algorithm, (i.e., data which belongs to a definite cluster could not be included in another cluster). Although it is the most favourable technique, it does have some weaknesses [13], [33]: 1. It is dependent on initialization. 2. It is sensitive to outliers and skewed distributions. 3. It may converge to a local minimum. 4. It may miss a small cluster. As a result, it may lead to poor or wrong representation of data. There are several clustering algorithms proposed to overcome the aforementioned weaknesses. Fuzzy C-means (FCM), an overlapping clustering that employs yet another fuzzy concept, allows each data to belong to two or more clusters at different degrees of memberships. In the FCM, there is no clear, significant boundary between the elements if they do, or do not belong to a certain class. In 2002, [34] successfully proposed a modified version of K-means clustering, namely, Moving K-Means (MKM) clustering. The study proved that MKM possesses a great ability in overcoming common problems in clustering, such as dead centres and centre redundancy. Furthermore, the MKM was proven to be effective in avoiding the centre from being trapped in local minima. A number of studies have also provided evidence that the MKM produced better performance as compared to the conventional KM and FCM [26], [27]. In this paper, we introduce a new version of clustering algorithm called Adaptive Fuzzy-K-means (AFKM) clustering algorithm. As mentioned, clustering is the process of organizing objects into groups wherein members are similar on certain aspects. In most clustering models, the concept of similarity is based on distances, such as the Euclidean distance. Agarwal and Mustaffa claimed that simply looking at Contributed Paper Manuscript received 10/08/10 Current version published 12/23/10 Electronic version published 12/30/10. 0098 3063/10/$20.00 © 2010 IEEE