In this paper we present an efficient geometrical method based on feature points derived from the skeleton of the normalized image. The type of these points and their relative position in the bounding-rectangle are used to get a code of 27 letters. Noise due to the skew, printing processes and quantization noise are examples of noise sources that may affect the character skeleton, distorting the feature points position. Experiments data with a presence of a noise of two pixels on the outline show that feature-points position can be modified about two pixels or one pixel in horizontal and vertical direction. We can therefore state that, in the case of Berber characters, feature points convey significant informations and this approach proved to be very tolerant for a certain degree of variation.