Automatic Image Annotation with Cooperation of Concept-Specific and Universal Visual Vocabularies

This paper proposes an automatic image annotation method based on concept-specific image representation and discriminative learning. Firstly, the concept-specific visual vocabularies are generated by assuming that localized features from the images with a specific concept are of the distribution of Gaussian Mixture Model (GMM). Each component in the GMM is taken as a visual token of the concept. The visual tokens of all the concepts are clustered to obtain a universal token set. Secondly, the image is represented as a concept-specific feature vector by computing the average posterior probabilities of being each universal visual token for all the localized features and assigning it to corresponding concept-specific visual tokens. Thus the feature vector for an image varies with different concepts. Finally, we implement image annotation and retrieval under a discriminative learning framework of Bayesian classifiers, Max-Min posterior Pseudo-probabilities (MM P). The proposed method were evaluated on the popular Corel-5K database. The experimental results with comparisons to state-of-the-art show that our method is promising.