A new method for combining dichotomizers like SVMs is proposed for classifying multi-class pattern fields. The novelty lies in the estimation of the styleconstrained posterior field class probabilities from the frequencies of the training patterns in the regions of the feature space engendered by the pairwise decision boundaries of the dichotomizers. We show that on simulated data, this non-parametric field classifier is nearly optimal. On scanned printed digits, its accuracy is comparable to that of state-of-the-art style classifiers.