A Semantic Context Model for Automatic Image Annotation

How to retrieve a required image from tens of thousands of images is challenging task. In this paper, we proposed a novel semantic context model for automatic annotation with context and spatial information. We reconstructed the image annotation as a multi-class classification problem and assign each object a label considering each object as an individual in both learning and annotation stage. And then, the class distribution of query region is estimated using Gaussian mixture model whose parameters are learned by expectation maximum algorithm. The posterior probabilities of all the concepts are obtained according to modified Bayesian rule. In the experiment, we conduct the performance evaluation on LabelMe image databases including 2651 images and 25653 regions. The experimental results illustrated that our proposed model effectively improve the performance of image annotation system, and the context information and height information could improve the precision of image annotation separately.