Scalable Markov model-based image annotation

In this paper, we propose a novel Markov model-based formulation for the image annotation problem. In this formulation, we treat image annotation as a graph ranking problem, by defining all possible labels in the lexicon as the states of a Markov chain. To fully utilize the correlation between labels, a query-biased transition matrix is dynamically constructed according to the query image. Based on this formulation, a scalable Markov model-based image annotation (MBIA) algorithm is presented to rank all the possible labels. To be scalable, we adapt search techniques on a Web-scale image set, and make MBIA capable of annotating arbitrary images with unlimited vocabulary. By fully exploring the correlation between labels, MBIA leads to superior performance than standard techniques. Experimental results on the typical Corel dataset and U. Washington dataset show the effectiveness and efficiency of the proposed algorithm.