Effective multi-modal multi-label learning for automatic image annotation

In this paper, we proposed a new multi-modal multi-label method for efficient automatic image annotation. Visual saliency analysis and multiple Nyström-approximating kernel discriminant analysis are adopted to obtain foreground semantic concepts. Region semantic analysis is used to get annotation words of background, and semantic correlation matrix by latent semantic analysis is used to improve the correctness of results. In our method, two different models are used to extract foreground and background annotation words respectively in terms of their distinct characters of semantic and visual features. Semantic correlation analysis could availably remove wrong labels for better results of multi-labeling. This approach has been evaluated on the Corel database, and compared with other algorithm. Experiment results show that our proposed method could achieve promising performance for multi-labeling, and outperform existing algorithm.