Improving automatic writer identification

State-of-the-art systems for automatic writer identification from handwritten text are based on two approaches: a statistical approach or a model-based approach. Both approaches have limitations. The main limitation of the statistical approach is that it relies on single-scale statistical features. The main limitation of the model-based approach is that the codebook generation is time-consuming. We attempt to improve automatic writer identification by overcoming the limitations of both approaches. For the statistical approach we evaluate multi-scale statistical features and find one of them to improve the identification performance of singlescale features. For the model-based approach we show that the usage of Kohonen maps for codebook generation is unnecessary, and that a randomly generated codebook is more efficient. We conclude that multi-scale features may enhance identification performances and that random codebooks are to be preferred over Kohonen-based codebooks.