Towards a ptolemaic model for OCR

In style-constrained classification often there are onlya few samples of each style and class, and the correspondencesbetween styles in the training set and the test setare unknown. To avoid gross misestimates of the classifierparameters it is therefore important to model the patterndistributions accurately. We offer empirical evidence for intuitivelyappealing assumptions, in feature spaces appropriatefor symbolic patterns, for (1) tetrahedral configurationsof class means that suggests linear style-adaptive classification,(2) improved estimates of classification boundariesby taking into account the asymmetric configuration of thepatterns with respect to the directions toward other classes,and (3) pattern-correlated style variability.