In the paper, we address a common issue of contaminated datasets, in which apart from proper native patterns we encounter garbage (foreign patterns). The required action is to remove foreign patterns. We propose a geometrical model for discrimination between native and foreign patterns. The idea is to enclose native patterns in elementary figures to discriminate between two areas: reserved for native and for foreign patterns. The choice of figures is dictated by the need for efficient computations. In the paper, we study a range of methods for constructing an optimal hyperrectangles-based model. The discussion covers empirical of handwritten digits recognition.