In this paper we presented a kernel density based Bayesian classifier for image object recognition. We used tangent distance to achieve tolerance with respect to affine transformations and proposed a distance measure tolerating small local variations called image distortion model. We related the two approaches and performed experiments on databases of different domains. Creating virtual training and test samples from the given data sets we obtained an excellent result of 2.2% error rate on the original USPS database. On a large database of radiographs both tangent distance and distortion model performed well and best results were obtained combining both approaches. Future work includes investigation of suitable transformations and cost functions for the generalized image distortion model (see section 4).