Bayesian Mixture Labeling by Minimizing Deviance of Classification Probabilities to Reference Labels