Abstract: Single feature is not discriminative to describe informational content of an image, which is always a shortcoming in traditional image retrieval task. It can be seen that one image can be described by different but complemented features. So multi-feature fusion ranking methods should be further researched to improve the ranking list of query in image retrieval. Most existing multi-feature fusion methods only focus on the nearest neighbors in image retrieval. When the ranking result of each neighbor graph is poor, it is hard to get ideal image retrieval result after graph fusion. In order to solve the problem, this paper proposes a novel multi-feature fusion method for image retrieval—GRF(group ranking fusion). The proposed method divides similar images of a data set into the same group, and improves the retrieval result of neighbor graph through similar images group. The GRF method expands the fusion scope in the premise of guaranteeing retrieval precision premise. At last, the experimental results on three standard data sets demonstrate that GRF can effectively utilize multi-feature graph to improve the performance of image retrieval.

Key words: multi-feature fusion, content-bases image retrieval, minimizing normalized cut, graph learning, retrieval re-rank