On image search result aggregation

In text retrieval, search result aggregation has been demonstrated how it has outperformed the retrieval results by single retrieval models. In general, search result aggregation for a specific query is based on combining different search results, which are produced by using different feature representations and/or different retrieval models. Particularly, several well-known combination methods, such as Borda count and CombSUM, have been proposed in the literature. However, in image retrieval the semantic gap problem limits the performances of current image retrieval systems. Since very few studies focus on search result aggregation in image retrieval, the aim of this paper is to assess the retrieval performances of different search result aggregation strategies and combination methods. Specifically, five different feature representations, five different distance functions as the retrieval models, and five different combination methods are used. Our experimental results based on Caltech 101, Caltech 256, and NUS-WIDE-LITE show that search result aggregation can definitely outperform single search results. In addition, among three aggregation strategies the one by combining five search results based on each best feature representation by their best distance function can provide the highest rate of precision rate.