An Image Annotation Method Based on CM Similarity Measure and Hybrid Relevance Feedback

Today, with the advent of digital imagery, the volume of digital images has been growing rapidly in different fields. So, there is an increasing need for providing an effective image retrieval system. In this paper, a semi-supervised k-means clustering method was introduced for image database clustering and image annotation. One of the most important parts of image clustering algorithms is to determine similarity of the images. To compute exact similarity measures, a new CM similarity measure was proposed here to make normalized and weighted features simultaneously so that similarity measure exploits normalized or weighted features in its formula to reach better performance. However, due to semantic gap, some images may be false clustered. A hybrid of three relevance feedback (RF) schemes was used to improve the accuracy of image clustering. (1) The images with the user who knows their irrelevance to a cluster were conducted to correct cluster by a long-term RF. (2) With regard to the images with the user who knows they are relevant to a cluster, feature weight of the clusters was estimated in order to provide a multiple similarity measure using a re-weighting RF. (3) To discover the exact place of the cluster centers, a cluster center movement (CCM) RF was used. Experimental results based on the Corel database including 1000 images and a satellite image database of Tehran city including 2400 images demonstrated the superiority of the proposed method in image database clustering.