A new method allowing to describe shapes from a set of polygonal curves using a relational descriptor is proposed in this paper. An approach based on discrete lines at several increasing widths is run on the contour of an object to provide a multi-level polygonal representation from accurate description to more and more rough aspects. On each polygon is calculated a force histogram to define a relational feature signature following a set of directions integrating both spatial relation organization and disparities of the shape in a same distribution. Three different matching schemes are proposed to compare multilevel distributions: global representation, level to level following extracted maxima. This new method is fast and a first experimental study achieved on a common database shows its good behavior.