Rotationally invariant moments constitute important techniques applicable to a versatile number of pattern recognition applications. Although the moments are invariant with regard to spatial transformations, in practice, due to the finite screen resolution, the spatial transformation themselves affect the invariance. This phenomenon jeopardizes the quality of pattern recognition. Therefore, this paper presents an experimental analysis of the accuracy and efficiency of discrimination under the impact of the most important spatial transformations such as rotation and scaling. We evaluate experimentally the impact of the noise induced by the spatial transformations on the most popular basis functions such as Zernike polynomials, Mellin polynomials and wavelets. The analysis reveals that the wavelet based moment invariants constitute one of the best choices to construct noise resistant features.