An empirical study of automatic image annotation through Multi-Instance Multi-Label Learning

Although many region based models for image auto-annotation have been proposed recently, their performances are not satisfactory due to the sensitivity to segmentation errors. In this paper, by evaluating two image partition methods and four visual features, we propose a new ensemble method under Multi-Instance Multi-Label (MIML) learning framework which has been proposed recently. The ensemble method combines all the outputs of these separate learning machines trained on different features. The experimental results over Corel images show that the ensemble method is efficient for image auto-annotation and comparable with other methods. In addition, the results show that the region-based image segmentation approach significantly improves the performance of the proposed model.