The recognition of primitives in digital geometry is deeply linked with separability problems. This framework leads us to consider the following problem of pattern recognition : Given a finite lattice set S ⊂ Z^d and a positive integer n, is it possible to separate S from Z d \S by n half-spaces? In other words, does there exist a polyhedron P defined by at most n half-spaces satisfying P ∩Z^d = S? The difficulty comes from the infinite number of constraints generated by all the points of Z^d \ S. It makes the decidability of the problem non straightforward since the classical algorithms of polyhe-dral separability can not be applied in this framework. We conjecture that the problem is nevertheless de-cidable and prove it under some assumptions: in arbitrary dimension, if the interior of the convex hull of S contains at least one lattice point or if the dimension d is 2 or if the dimension d = 3 and S is not in a specific configuration of lattice width 0 or 1. The proof strategy is to reduce the set of outliers Z^d \ S to its minimal elements according to a partial order "is in the shadow of ". These minimal elements are called the lattice jewels of S. We prove that under some assumptions, the set S admits only a finite number of lattice jewels. The result about the decidability of the problem is a corollary of this fundamental property.