Visualisation, assembling/disassembling simulation, reverse engineering, or finite element analyses are so many applications where a polyhedral simplification can be mandatory. In most of the configurations, such an operation must preserve the significant shapes of the original model. This paper addresses the way these simplifications can be performed while preserving character lines, i.e. those lines which affect significantly the object visual appearance. The proposed simplification process is based on a vertex removal algorithm that uses a heterogeneous map of sizes computed with information extracted from 2D images of the object. The farther the vertices are from the character lines, the more the simplification is important. The geodesic distances are estimated while using a new algorithm based on the analysis of multiple filtering of the images. This method might reduce the computational time since there is no need to compute either the curvature or the saliency. Thus, a first attempt to enrich a polyhedral model with semantic information extracted from images is proposed.