Image Re-ranking using Information Gain and Relative Consistency through Multi-graph Learning

After receiving a lot of attention towards text based searching for image retrieval, researchers have focused on content based image retrieval. Visual re-ranking is a method of image retrieval, which has been widely accepted to boost the accuracy of traditional text-based image retrieval. Current trend of this method is to combine the retrieval results from various visual features to boost the overall performance. The challenge in this trend of re-ranking is to exploit the complementary property of different features effectively. Our purpose basically comes under feature based image retrieval on three different modalities, so that retrieval re-ranking will be more accurate and effective. We deal with mainly two terms: information gain and relative ranking consistency among multiple modalities. Our submodular re-ranking framework can be easily used in re-ranking problems for realtime search engines.