Image Search Reranking with Relevance, Diversity and Topic Coverage
Xuefei Lin, Tian Zhang · 2016
Image search reranking has recently been proposed to improve image search results. Most of the conventional reranking methods cannot leverage both relevance and diversity of the search results simultaneously. In addition, they usually ignore the latent topics of images. Towards this end, this paper proposes a new reranking method by exploring relevance, diversity and topic coverage of the search results simultaneously. Specifically, the proposed method first groups the returned images by exploring the underlying topics. Then, the desired ranking list is generated by a greedy algorithm based on the relevance score, visual similarity, topic coverage and representativeness scores. Experimental results on an image set collected from Flickr demonstrate the performance of the proposed method.