Automatic Image Annotation and Refinement Based on CCA Subspace

In the field of automatic image annotation, researchers rarely considered the relations between low-level features of images and tags, and also rarely takes into account that describing the probability distributions of keywords from the only side of tags. For this reason, this paper proposed a method called automatic image annotation and refinement based on canonical correlation analysis (CCA) subspace. The method estimates the correlation features between image features and tags text using CCA, then projects the features of testing images to the CCA subspace, then the image will be annotated according to the projected image features. The class density probability of every word is estimated using Gaussian mixture model (GMM). The probabilities of the tags with respect to the image are calculated using Bayesian classifier. In the end, the candidate annotations can be refined and re-ranked using the semantic correlation between tags. This method accurately built a “bridge” between low-level features of image and semantic features. Experiments results demonstrate that the proposed method has significantly improved the precision and recall.