Tag Quality Improvement for Social Image Hosting Website

Abstract

Social image hosting websites such as Flickr provide services to users for sharing their images. Users can upload and tag their images or search for images by using keywords which describe image semantics. However various low quality tags in the user generated folksonomy tags have negative influences on image search results and user experience. To improve tag quality, we propose four approaches with one framework to automatically generate new tags, and rank the new tags as well as the existing raw tags, for both untagged and tagged images. The approaches utilize and integrate both textual and visual information, and analyze intra- and inter- probabilistic relationships among images and tags based on a graph model. The experiments based on the dataset constructed from Flickr illustrate the effectiveness and efficiency of our approaches.