Skeletonization is a widely used transformation in the elds of image analysis and shape recognition. The concept
of skeleton was introduced by Blum in 1961 [4], under the name of medial axis transform. The denition proposed
by Blum corresponds to the intuitive skeleton notion, a sort of minimal representation of a set X under the form of
lines of unit width. It is based on the concept of prairie re: assuming a re is propagating within X at a uniform
speed starting from its contours, the medial axis of X is the set of the points where dierent refronts meet (see
Fig. 1). A more formal denition of skeletons in both the discrete and the continuous cases was proposed by Calabi
in 1966 [6] and is based on the concept of maximal bal l. This denition tends to be more widely used in the eld of
morphological image analysis [17], and this is the reason why we recall it in x 1.1. In this paper, we consider both
approaches.
The topological study of the skeleton in the continuous plane IR2
turns out to be extremely complex. One of the
ma jor results states that the topological closure of the skeleton of a connected open set is itself connected [10, 11]. . .
Similarly, in the discrete case, the skeleton notion is far from being a trivial one. Indeed, as we shall see later,
the direct application of Calabi's denition leads to skeletons which do not preserve the homotopy of the sets on
which they act. Various algorithms have been proposed to remedy this inconvenience, none of which is completely
satisfying: they either produce skeletons with parasitic dendrites, or are computationally inecient, or only allow
the computation of one given skeleton-like transformation. In fact, the literature on skeletons is so abundant that
it becomes impossible to review all the existing works exhaustively. We shall just brie
y mention below the main
categories of skeleton algorithms, so as to show in what respects the present approach is novel.