Spectral Representations of Linear Features for Generalisation

In this paper we propose the use of new representations of linear features in order to make up for the weakness of classical generalisation algorithms. Such representations are developed from spectral tools: Fourier series and wavelet decomposition. The theoretical possibilities of representations are discussed from a generalisation point of view. Algorithms are built on these representations. The experiments are presented and discussed, particularly the encouraging results leading to caricature.