A project on topographic maps interpretation is being driven at I.G.N (Institut Geographique National, France). The aim of this project is to provide a totally automatic interpretation with a call to operator on the only suspicious cases detected by the program. Main difficulties of this task are due to maps' density and variety of legend. To overcome this problem, we have adopted a peeling onion strategy where we progressively subtract of the map the already recognized layers to facilitate the interpretation of remaining more complex layers. We present numerous results of the layers already extracted. The most important point to address in the near future, will be the computation of reliability indicator allowing to throw an alert-process.