Image tagging plays a critical role in image indexing and retrieval and it has gained more and more attention along with the increasing availability of large quantities of web images. However, most of current tagging methods only utilize single feature type, while combining multiple types of features has been proved to be effective for image analysis. In this paper, we propose a multi-feature late fusion method for image tagging. For an image, we first learn several scores with regard to each tag by using different single features or combinations of single features based on a tag relevance learner. Then we learn an optimal combination weight for each tag score and linearly combine all the tag scores with the learned weights. Finally, a low-rank tag pairwise matrix is learned with the linearly combined tag scores and a robust tag score is recovered from the low-rank matrix. The tags with the largest scores are regarded as the predicted tags. We compare our approach with several multi-feature fusion techniques over a real-world dataset NUS-WIDE and show the effectiveness of the proposed multi-feature fusion method.
