Image Annotation combining Subspace Clustering , Matrix Completion and Inhomogeneous Errors

Image annotation methods have greatly facilitated image management applications. However, existing methods are still suffering from the degradation of the missing and noisy tags provided by users. In this study, we propose an image annotation method which performs tag completion and refinement sequentially. We assume that images are sampled from a union of subspaces. Images sampled from the same subspace, as well as their corresponding tags, should form a compatible image-tag sub-matrix. Thus we segment the subspaces by the Sparse Subspace Clustering (SSC) method and share tags in each subspace. A novel matrix completion model is designed for tag refinement, taking visual-tag correlation, semantic-tag correlation and the inhomogeneous errors property, which is explored in this field for the first time, into consideration. We exploit CNN features to improve the model. The proposed algorithm outperforms state-of-the-art approaches when handling missing and noisy tags on multiple benchmark datasets.