Proceedings PaperSelection of image features for distribution-map classifiers
|Format||Member Price||Non-Member Price|
A useful metric for pattern classification can be derived from a series of one-dimensional maps of empirical class-conditional distributions. To achieve maximum classification accuracy, the distributions need to be projected into subspaces where they are locally dense and the classes are at least partially separable. We describe a method where useful projections are obtained automatically by a sequence of cluster and linear discriminant analyses. Each projection can be considered as a feature that contributes to the elimination of unresolved ambiguities. We discuss conditions under which the method can achieve maximum accuracy. In an experiment of applying the method to machine-printed character images, we show that the method yields a classifier that has very low error rate, and can be improved with additional training data.