Pattern Classification Based on a Piecewise Multi-linear Model for the Class Probability Densities
When a Bayesian classifier is designed, a model for the class probability density functions (PDFs) has to be chosen. This choice is determined by a trade-off between robustness and low complexity — which is usually satisfied by simple parametric models, based on a restricted number of parameters — and the model’s ability to fit a large class of PDFs — which usually requires a high number of model parameters.
In this paper, a model is introduced, where the class PDFs are approximated as piecewise multi-linear functions (a generalisation of bilinear functions for an arbitrary dimensionality). This model is compared with classical parametric and non-parametric models, from a point of view of versatility, robustness and complexity. The results of classification and PDF estimation experiments are discussed.
KeywordsQuadratic Discriminant Analysis Orthogonal Expansion Localise Base Function Orthonormal Base Function Simple Parametric Model
- 1.Richard O. Duda and Peter E. Hart. Pattern Classification and Scene Analysis. John Wiley & Sons, 1973.Google Scholar
- 2.Julius T. Tou and Raphael C. Gonzales. Pattern Recognition Principles. Addison Wesley Publishing Company, 1979.Google Scholar
- 3.Robert Schalkoff. Pattern Recognition — Statistical, Structural and Neural Approaches. John Wiley & Sons, 1992.Google Scholar
- 4.Fang Sun, Shin’ichiro Omachi, and Hirotomo Aso. An algorithm for estimating mixture distributions of high dimensional vectors and its application to character recognition. In Proc. 11th Scandinavian Conference on Image Analysis, pages 267–274, 1999.Google Scholar
- 6.N. Balakrishnan. Introduction and historical remarks. In N. Balakrishnan, editor, Handbook of the Logistic Distribution. Marcel Dekker, Inc, 1992.Google Scholar
- 7.W. J. Dixon, M. B. Brown, L. Engelman, J. W. Frane, M. A. Hill, R. I. Jennrich, and J. D. Toporek. BMDP Statistical Software 1981, 1981.Google Scholar