Feature reduction and selection for automatic image annotation

Abstract

Automatic image annotation for large collections of images is a challenging problem. For labeling images precisely, more various features including low-level image features, EXIFs, textual tags of images are expected to be used. However, not all features contribute useful information for each concept. The high-dimension problem causing by combining all features is detrimental to the concept learning. In this paper we propose the feature reduction and selection method to improve the performance of annotating images. The proposed feature reduction methods extract informative features to reduce the dimensions. While the feature selection method based on the wrapper model can select effective features from miscellaneous features. The experimental result shows that the proposed feature reduction method improves the efficiency of concept learning. The developed feature selection method also increases the labeling precision and recall of images. © Springer-Verlag Berlin Heidelberg 2013.