The paper addresses the problem of pattern classification when distortions are present in the observed data. These are distortions occurring either in the process of classifying a test pattern or during the learning phase of a classifier design. In particular we consider the case when the measurements are corrupted by noise. Additive, multiplicative and generalized noise models are taken into account. A unified framework of the posed problems based on the Bayesian paradigm is given. Learning algorithms stemming from nonparametric curve estimation techniques are derived. We examine the effect of distortions in measurements on the proposed classification algorithms. A class of classification techniques being Bayes risk consistent is derived. The latter result makes use of the idea of deconvolution.