Aggregation of Multiple Pseudo Relevance Feedbacks for Image Search Re-Ranking

Image retrieval effectiveness can be improved by pseudo relevance feedback (PRF), which automatically uses top-k images of the initial retrieval result as the pseudo feedback. Since there are several different strategies for performing PRF leading to different search results, in this paper we focus on image search re-ranking by search result aggregation as a hybrid approach. In particular, different combinations of the original retrieval result with the result of PRF and the result of PRF by pseudo positive and negative feedbacks, using a strategy based on the Borda count, are compared. Our experiments, carried out on the NUS-WIDE-LIT and Caltech 256 datasets, demonstrate that search result aggregation can provide better retrieval performance than PRF. Specifically, the combination of the original result and the result of PRF by pseudo positive feedback performs the best.