A METHOD OF CLUSTERING FOR CONTENT-BASED IMAGE RETRIEVAL

Abstract

In this paper, an improvement in K-Means algorithm was proposed to cluster and applied to the problem of searching similar images by content. To accomplish this, we used a threshold value that measured the similarity between data objects, which is called as θ. K-Means algorithm was improved by not pre-determining the number of cluster centers, the number of data clusters grow with the increase in the number of images. The image was extracted as a n-dimensional vector and was an input for the improved K-Means algorithm from which to search for similar images. In order to demonstrate the proposals, we experimented and evaluated the results on the COREL image data set (1000 images) and compared to other recently published works on the same dataset. According to the experimental results, our proposals are feasible and applicable to different image retrieval systems.