Image annotation by modeling Supporting Region Graph

Annotating image regions with keywords has received increasing attention in the computer vision community in recent years. Recent studies have shown that graphical modeling techniques, such as Conditional Random Fields (CRF), greatly improves the accuracy of image annotation by utilizing contextual information among image regions. However, training and predicting with the high-order CRF is computational expensive so that only adjacent regions can be utilized to build its graph structure. In this paper, we develop a light-weight classification model, Approximated Supporting Region Graph (ASRG), in order to handle more relevant regions efficiently, with which a large number of supporting regions are selected and their features are utilized to represent the contextual information in the training and prediction for each image region. Experimental results show that our model is much more computational efficient and achieves competitive performance comparing with CRF and other state-of-art methods.