The underlying purpose of this thesis is to investigate methods of fingerprint recognition which employ principles from the field of machine learning, and that do not require much image pre-processing. Fingerprint images are represented by features derived from their spectrum. The features are to a certain extent invariant with respect to translation and rotation. The features are chosen such that the two classes of identical and different fingerprints are best separated. The features are used in different classification methods, namely nearest neighbour classifier, MLP and SVM, with specifically adapted training methods. The resulting matchers are compared. In addition, a pre-matcher based on these features is constructed, together with figures to describe its characteristics. This thesis makes two major contributions: The spectral features are optimised such that a low-dimensional representation with high discriminative power is obtained. Secondly, the proposed pre-matching system drastically reduces both the error rates and the overall runtime, while the pre-matcher itself requires only little computational effort.