Web image annotation using two-step filtering on social tags

Web image annotation has become an important issue with exploding web images and the necessity of effective image search. The social tags have recently utilized at image annotation because they can reflect the user's tagging tendency, and reduce the semantic gap. However, an effective filtering procedure is required to extract the relevant tags since the user's subjectivity and noisy tags. In this paper, we propose a two-step filtering on social tags for image annotation. This method conducts the filtering and verification tasks by analyzing the tags of visual neighbor images using voting method and co-occurrence analysis. Our method consists of the following three steps: 1) the tag candidate set is founded by searching the visual neighbor images, 2) from a given tag candidate set, coarse filtering is conducted by tag grouping and voting technique, 3) the dense filtering is conducted by using similarity verification for coarse filtered candidate tag set. To evaluate the performance of our approach, we conduct the experiments on a social-tagged image dataset obtained from Flickr. We compare the annotation accuracy between the voting method and our proposed method. Our experimental results show that our method has an improvement in image annotation.