Abstract

Image searching is a key issue in nowadays social community service. However, for a given query the search results are usually noisy and lacking visual diversity. In order to solve this issue we propose a novel approach to re-rank the initial search results. We first build two graphs based on the visual and textual similarities among images. Then perform random walks on the two graphs. We use two strategies to fuse the results derived from random walks to leverage images contents and users annotations. Finally, we utilize a penalty process to enhance the visual diversity of the output. Experimental results demonstrate the effectiveness of the presented model.

Paper Details