Region-based annotation of digital photographs

Abstract

We propose a region-based method for the annotation of outdoor photographs. First, images are oversegmented using the normalized cut algorithm. Each resulting region is described by color and texture features, and is then classified by a multi-class Support Vector Machine into seven classes: sky, vegetation, snow, water, ground, street, and sand. Finally, a rejection option is applied to discard those regions for which the classifier is not confident enough. For training and evaluation we used more than 12,000 images taken from the LabelMe project. © 2011 Springer-Verlag Berlin Heidelberg.