A novel algorithm for filtering out junk images interactively from web search results

Abstract

Keyword-based web search engine uses text to reflect users' query intentions. However, it is hard to descript user's intention with simple text terms accurately, and besides of this, it is also hard to make the association between the text terms and images precisely. As a result, the keyword-based image search engine may return large amount of junk images. In this paper, an interactive image filter algorithm is proposed, which utilizes the user's search intention for further filtering out junk images from web search result. The returned images of web search are divided into groups with a multiple kernels image clustering technique firstly. Then, the hyperbolic visualization is adopted to display these images for users to assess the relevance between the query intentions and search results, and the junk images are filtered out interactively. The above steps repeat until the user's requirement is reached. Experiments on diverse queries result show that the proposed method can improve the engine's precision rate effectively while effectively control the false detection rate.