In this paper, the Radon transform method is used to generate a set of rotation invariant characteristics. Experiments of our approach were carried out on a database of ten decimal digits (0 to 9) in 24 different orientations from 0° to 360 ° by a step of 15 °. A multilayer perceptron neural network is used in the classification phase to test the effectiveness of our approach. The proposed approach is noise-effective and leads to a classification rate equal to 100 % for images without noise and a classification rate equal to 95.2 for images with noise.