Annotating images with tags is useful for indexing and retrieving images.However, many available annotation data include missing or inaccurateannotations. In this paper, we propose an image annotation framework whichsequentially performs tag completion and refinement. We utilize the subspaceproperty of data via sparse subspace clustering for tag completion. Then wepropose a novel matrix completion model for tag refinement, integrating visualcorrelation, semantic correlation and the novelly studied property of complexerrors. The proposed method outperforms the state-of-the-art approaches onmultiple benchmark datasets even when they contain certain levels of annotationnoise.
展开▼