Time-frequency dictionaries for improved discriminant feature extraction

Reduction of signal dimensionality in the pre-classification stage of classification systems is usually done via one of many classical parameter extraction methods, for example linear predictive modelling or Fourier analysis. Many of these methods concentrate on the best possible signal representation and help the subsequent classification stage only in that they have effectively compressed (according to a particular criterion) the signal thus requiring less training samples. In this paper, we consider the situation where the classification is helped in the feature extraction stage by supplying it with good discriminative features obtained by transformation onto a coordinate system consisting of a collection of orthonormal functions that are well localised in both time and frequency and then choosing the most suitable of these for the problem at hand.