A two-stage hybrid probabilistic topic model for refining image annotation

Refining image annotation has become one of the core research topics in computer vision and pattern recognition due to its great potentials in image retrieval. However, it is still in its infancy and is not sophisticated enough to extract perfect semantic concepts just according to the image low-level features. In this paper, we propose a two-stage hybrid probabilistic topic model to improve the quality of automatic image annotation. To start with, a probabilistic latent semantic analysis model with asymmetric modalities is learned to estimate the posterior probabilities of each annotation keyword, during which the image-to-word relation can be well established. Next, a label similarity graph is constructed by a weighted linear combination of label similarity and visual similarity of images associated with the corresponding labels. By this way, the information from image low-level visual features and high-level semantic concepts can be seamlessly integrated by fully taking into account the word-to-word and image-to-image relations. Finally, the rank-two relaxation heuristics is exploited to further mine the correlation of the candidate annotations so as to capture the refining results, which plays a critical role in semantic based image retrieval. Extensive experiments show that the proposed model achieves not only superior annotation accuracy but also better retrieval performance.