A comparison of image grouping techniques of content based image retrieval using K-means clustering algorithm

Abstract

Content Based Image Retrieval (CBIR) system is an alternative approach to Text Based Image Retrieval (TBIR) system in retrieving the images. The system consists of three phases which are feature extraction, image grouping and image retrieval. This study focused on colour feature for feature extraction process, image grouping for grouping images according to their characteristic similarities. For image retrieval, several well-known clustering techniques were introduced and applied to CBIR system. The clustering technique of K-Means type is the most preferable clustering technique since it is easy to be implemented and also fast computation. However, because of many improvement that have been done towards this technique, there exist variations of K-Means clustering algorithms. Thus, in this research, a comparison performance among three types of K-Means clustering algorithms, namely the basic K-Means, Fuzzy K-Means and K-Harmonic Means algorithms is performed. Four validation techniques are used for determining the most efficient algorithm in retrieving the images, which were Davies-Bouldin index (DB), Calinski-Harabasz index (CH), Dunn index (Dunn) and Silhouette width (SC). Based on these four validation techniques, the K-Harmonic Means clustering algorithm was found to be the best clustering algorithms in grouping image dataset.