Image retrieval via generalized I-divergence in the bag-of-visual-words framework

The focus of this work is on the use of Generalized I-Divergence (GID) applied to content-based image retrieval from a real scenes database using the Bag-of-Visual-Words (BoVW) in three different approaches: BoVW, BoVW with the Binary Saliency Map (BSM) and BoVW with the Fuzzy Image Saliency Map (FISM). The BoVW approach has shown a promising performance within the context of the image retrieval, where in general, the similarity between two images is typically measured by the distance between the two histograms using the cosine similarity measure. In this work we use Divergence measures to rank similar images instead of cosine measure with the aim of improving the performance of the methods. Experimental results from two real world image databases have shown that our approach achieved the best precision rates when compared to the use of cosine similarity, particularly on queries with poor initial recall where the traditional cosine distance fails to measure the similarity between the hyperplanes which represent the query image features and the database image features.