A Discriminant Function for Noisy Pattern Recognition
Shin'ichiro Omachi, Fang Sun, and Hirotomo Aso
Proceedings of The 11th Scandinavian Conference on Image Analysis (SCIA'99), pp.793-800, June 1999

Abstract
For a practical pattern recognition system, noisy pattern recognition is necessary and important. There are several basic ideas of recognizing noisy patterns, for example, constructing a dictionary with noisy patterns, applying different classifiers, or using filters to delete noise. In most conventional statistical pattern recognition methods, a feature vector is extracted from an object. The distribution of feature vectors is estimated for each category to select a candidate for an unknown input pattern. As it is known, the distribution of feature vectors will change if noise occurs. It is impossible to predict all kinds of noise that happen accidentally and irregularly. In this paper, an attempt to deal with noisy patterns is examined, and a new discriminant function is proposed. By introducing a revision matrix calculated from an unknown input pattern, the new function can be defined as a revised form of traditional function with the degree of detected noise of each individual pattern. The ability of the proposed function is evaluated by experiments on discrimination of two categories. Experimental results not only show the classification effectiveness of the proposed function, but also confirm that it is enormously important to detect the noise and to revise the discriminant function for noisy pattern recognition.
Full paper
PDF
Gzipped Postscript