A hybrid hierarchical framework for automatic image annotation

Automatic image annotation is a challenging problem in multimedia content analysis and computer vision. In this paper, we propose a hierarchical framework for multi-label image annotation. We first present an image-filtering algorithm to remove most of the irrelevant images for an unlabeled image. In the image-filtering algorithm, an image cluster is allocated using a discriminative model as the primary relevant image set for the unlabeled image. And then the relevant images are updated by making use of the relationships between semantic concepts. In the next stage, a hybrid annotation model is proposed to annotate images. On one hand, we present a baseline method to transfer labels from relevant images to unlabeled image according to global visual features. On the other hand, regional visual features are extracted to build a probabilistic model for image annotation. Finally, the two annotation results are fused by a simple weighted algorithm. Experiments have proved that our hierarchical framework outperformed the current state-of-the art models for image annotation.