MFS-Map: efficient context and content combination to annotate images

Abstract

Automatic image annotation provides textual description to images based on content and context information. Since images may present large variability, image annotation methods often employ multiple extractors to represent visual contente considering local and global features under different visual aspects. As result, an important aspect of image annotation is the combination of context and content representations. This paper proposes MFS-Map (Multi-Feature Space Map), a novel image annotation method that manages the problem of combining multiple content and contexto representations when annotating images. The advantage of MFS-Map is that it does not represent visual and textual features by a single large feature vector. Rather, MFS-Map divides the problem into feature subspaces. This approach allows MFS-Map to improve its accuracy by identifying the features relevant for each annotation. We evaluated MFSMap using two publicly available datasets: MIR Flickr and Image CLEF 2011. MFS-Map obtained both superior precision and faster speed when compared to other widely employed annotation methods.