Fast Image Auto-annotation with Visual Vector Approximation Clusters

The paper proposes a simple novel technique to automatically determine a set of keywords that describe the content of an image. The images are segmented in ‘blobs’, which are approximatively classified using discretized features space. This results in a small number of visual Vector Approximation Clusters (VAC), which allows to train the joint probability table of the visual features and the textual annotations from a training data set. Futhermore a simple Bayes model is used to determine the probability that a keyword describes a test image. The paper includes an experimental evaluation on COREL database. We compare our approach with state of the art auto-annotation methods using the same database, words set and scoring method. Results show that our simple method give similar results than state of the art models.