A Hypergraph-Based Reranking Model For Retrieving Diverse Social Images

With the proliferation of social networks and photo-sharing websites, the need for an effective image retrieval system has become crucial. To match the users' intents, retrieval results are expected to be not only relevant to the query but also diverse. In this way, they depict a comprehensive summarization of the user query. Motivated by this observation, we propose a hypergraph-based reranking model for retrieving diverse social images. Indeed, a visual hypergraph is constructed to capture high-order relationships among images. Different from exiting hypergraph ranking that usually ranks images according to their relevance to a given query, our approach emphasizes diversity by integrating absorbing nodes into the ranking process. This way, redundant images are prevented from getting high ranking scores, thereby ensuring diversity. Extensive experiments conducted on the MediaEval 2016 dataset demonstrate that our approach can achieve competitive performance to the existing diversification approaches.