Local and global approaches for unsupervised image annotation

Image annotation is the task of assigning keywords to images with the goal of facilitating their organization and accessibility options (e.g., searching by keywords). Traditional annotation methods are based on supervised learning. Although being very effective, these methods require of large amounts of manually labeled images, and are limited in the sense that images can only be labeled with concepts seen during the training phase. Unsupervised automatic image annotation (UAIA) methods, on the other hand, neglect strongly-labeled images and instead rely on huge collections of unstructured text containing images for the annotation. In addition to not requiring labeled images, unsupervised techniques are advantageous because they can assign (virtually) any concept to an image. Despite these benefits, unsupervised methods have not been widely studied in image annotation, a reason for this is the lack of a reference framework for UAIA. In this line, this paper introduces two effective methods for UAIA in the context of a common framework inspired in the way a query is expanded throughout Automatic Query Expansion (AQE) in information retrieval. On the one hand, we describe a local method that processes text information associated to images retrieved when using the image to annotate as query, several methods from the state of the art can be described under this formulation. On the other hand, we propose a global method that pre-process offline the reference collection to identify visual-textual associations that are later used for annotation. Both methods are extensively evaluated in benchmarks for large-scale UAIA. Experimental results show the competitiveness of both strategies when compared to the state of the art. We foresee the AQE-based framework will pave the way for the development of alternative and effective methods for UAIA.