This paper is addressed to the problem of typography extraction in document images. For that, we propose the use of robust textural image processing methods (Gabor filtering and the geodesic active regions model) instead of the classical document image processing techniques (physical segmentation and logical labeling) which work at a pixel level and are very sensitive to a lot of parameters such as: noise, skewing... We show, on a few examples, that our method is generic enough to cope with a lot of recurrent problems in the field of document processing.