Object classification using the MPEG-7 visual descriptors: an experimental evaluation using state of the art data classifiers

Abstract

MPEG-7 visual descriptors include the color, texture and shape descriptor and were introduced, after a long period of evaluation, for efficient content-based image retrieval. A total of 22 different kind of features are included, nine for color, eight for texture and five for shape. Encoded values of these features vary significantly and their combination, as a means for better retrieval, is neither straightforward nor efficient. Despite their extensive usage MPEG-7 visual descriptors have never compared concerning their retrieval performance; thus the question which descriptor to use for a particular image retrieval scenario stills unanswered. In this paper we report the results of an extended experimental study on the efficiency of the various MPEG-7 visual features with the aid of the Weka tool and a variety of well-known data classifiers. Our data consist of 1952 images from the athletics domain, containing 7686 manually annotated objects corresponding to eight different classes. The results indicate that combination of selected MPEG-7 visual features may lead to increased retrieval performance compared to single descriptors but this is not a general fact. Furthermore, although the models created using alternative training schemes have similar performance libSVM is by far more effective in model creation in terms of training time and robustness to parameter variation