A Novel Image Segmentation Algorithm Based on Harmony Fuzzy Search Algorithm Osama Moh’d Alia, Rajeswari Mandava, Dhanesh Ramachandram CVRG, School of Computer Sciences Universiti Sains Malaysia 11800 USM, Penang, Malaysia sm alia@yahoo.com, {mandava,dhaneshr}@cs.usm.my Mohd Ezane Aziz Department of Radiology, Health Campus Universiti Sains Malaysia 11650 K.K, Kelantan, Malaysia drezane@cs.usm.my Abstract—Image segmentation is considered as one of the cru- cial steps in image analysis process and it is the most challenging task. Image segmentation can be modeled as a clustering problem. Therefore, clustering algorithms have been applied successfully in image segmentation problems. Fuzzy c-mean (FCM) algorithm is considered as one of the most popular clustering algorithm. Even that, FCM can generate a local optimal solution. In this paper we propose a novel Harmony Fuzzy Image Segmentation Algorithm (HFISA) which is based on Harmony Search (HS) algorithm. A model of HS which uses fuzzy memberships of image pixels to a predefined number of clusters as decision variables, rather than centroids of clusters, is implemented to achieve better image segmentation results and at the same time, avoid local optima problem. The proposed algorithm is applied onto six different types of images. The experiment results show the efficiency of the proposed algorithm compared to the fuzzy c-means algorithm. Index Terms—harmony search algorithm; image segmentation; fuzzy clustering; cluster validity index. I. I NTRODUCTION Image segmentation is the process of subdividing the digital image into their constituent regions, in which each region shares similar properties or features [1]. These features can be brightness, spatial coherence, color, texture, motion, mean, variance, etc. The level of subdivision depends on the problem to be solved. Since no single image segmentation algorithm can solve all types of image segmentation problems, many algorithms have been researched and presented, each of which uses a different induction principle. From the literature, many classifications were assigned to segmentation algorithms, such as thresholding methods, deformable models, clustering meth- ods, histogram based methods, region based methods, graph partitioning methods, classification methods, and etc. [2], [3]. Image segmentation can be modeled as clustering problem [4]. Therefore, clustering algorithms have been applied suc- cessfully in image segmentation problems. They can be cate- gorized into different groups, such as hierarchical algorithms, partitional algorithms, density-based clustering algorithms, fuzzy clustering, etc [1]. Fuzzy c-means algorithm (FCM) [5] is considered as one of the most popular algorithm used in clustering problems and at same time in image segmentation problems [6], [7], [8]. However FCM algorithm has serious limitations such as the tendency to become trapped in local optima and prone to initialization sensitivity [9], [10]. One approach to obtain a globally optimum solution and overcome these limitations is to use one of the metaheuristic algorithms to find the appropriate initial cluster centers and then feed these cluster centers into FCM algorithm. In this approach, the conventional way is to consider cluster centers as an optimization decision variables and find the appropriate values for them using the abilities of metaheuristic algorithms to exploring and exploiting the search space such as genetic al- gorithms [11], simulated annealing [12], tabu search algorithm [13], bees colony algorithm [14], particle swarm optimization algorithm [15], ant colony algorithm [16], [17], and recently harmony search (HS) algorithm [18], [19], [20], [21]. In this paper, we propose a novel approach to image segmentation called Harmony Fuzzy Image Segmentation Al- gorithm (HFISA). It is based on Harmony Search (HS) algo- rithm, where HS is also a new metaheuristic population based algorithm which was developed by Geem [22] and has been successfully tailored to different computational optimization problems [23]. In this paper, a new approach to select the decision variables for HS is proposed. This is done by using fuzzy memberships of image pixels to the predefined number of clusters as a decision variables, rather than centroids of clusters in this optimization problem. The effectiveness of HFISA is demonstrated on six different types of images such as natural image, synthetic images, medical MR image and remote sensing image. The rest of the paper is organized as follows: Section II describes the fundamentals of fuzzy clustering with FCM. Section III describes fuzzy cluster validity measures. Section IV discusses the proposed HFISA. Section V shows the experimental results and the final section, presents conclusion and future directions of our work. II. FUNDAMENTALS OF FUZZY CLUSTERING Clustering is a typical unsupervised learning technique for grouping similar data points (pixels) according to some