Feature Transforms and Classification

Abstract

One single feature is usually not sufficient to detect, characterize or classify the content of a multimedia signal. Decisions typically have to be derived from multiple features; those can be quite different in terms of physical meaning, but also statistically may be significantly distinct in terms of value ranges, mean value, variance, etc. Further, linear or nonlinear statistical dependencies may exist between different features, which have to be taken into consideration for similarity criteria and classification. In this chapter, methods of feature weighting and transforms are introduced first, where the latter may map the set of feature values into a possibly small number of independent values. The classification itself requires similarity criteria by which a certain feature constellation can be matched against characteristic properties of a class, or by which similarity of features from two different signals is judged. Different similarity criteria are discussed, including the comparison of single feature values, statistical distributions and of data sets. Furthermore, criteria for judging the quality of classification are introduced. Various classification methods are discussed, where optimization typically relies on class definitions obtained by training sets, but some approaches also allow blind classification or partitioning of data sets. Linear classification, cluster-based and nearest neighbor methods, Bayes classification, artificial neural networks, support vector machines and boosting are some state of the art classification approaches that are explained in detail. Concepts of quantifying uncertainty and evidence are presented in the end of this chapter.