Distance measures for signal processing and pattern recognition

Abstract Some general tools for measuring distances either between two statistical models or between a parametric model (or signature) and a signal are presented. These tools are useful in a variety of signal processing applications such as detection, segmentation, classification, recognition and coding. After a section devoted to general distance measures between probability laws, the question of spectral distances between processes is investigated. Then results concerning AR and ARMA models are described. Problems related to the interaction between distances for parametric models and estimation of the parameters of these models are also mentioned. Also recalled (when necessary) are some classical results about error bounds in classification and feature selection for pattern recognition, which are obtained with the aid of properties of distance measures.