Feature Scaling in Support Vector Data Descriptions

Abstract

In previous research the Support Vector Data Description is proposed to solve the problem of One-Class classification. In One-Class classification one set of data, called the target set, has to be distinguished from the rest of the feature space. This description should be constructed such that objects not originating from the target set, by definition the outlier class, are not accepted by the data description. In this paper the Support Vector Data Description is applied to the problem of image database retrieval. The user selects an example image region as target class and resembling images from a database should be retrieved. This application shows some of the weaknesses of the SVDD, particularly the dependence on the scaling of the features. By rescaling features and combining several descriptions oll well scaled feature sets, performance can be significantly improved.