We consider the problem of automatic image tagging for online services and explore a prototype-based approach that applies ideas from manifold ranking. Since algorithms for ranking on graphs or manifolds often lack a way of dealing with out of sample data, they are of limited use for pattern recognition. In this paper, we therefore propose to consider diffusion processes over bipartite graphs which allow for a dual treatment of objects and features. As with Google's PageRank, this leads to Markov processes over the prototypes. In contrast to related methods, our model provides a Bayesian interpretation of the transition matrix and enables the ranking and consequently the classification of unknown entities. By design, the method is tailored to histogram features and we apply it to histogram-based color image analysis. Experiments with images downloaded from flickr.com illustrate object localization in realistic scenes.
展开▼
