# Discrete approximation of the Mumford-Shah functional in ... Free discontinuity problems, special...

date post

17-Jul-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Discrete approximation of the Mumford-Shah functional in ... Free discontinuity problems, special...

Mathematical Modelling and Numerical Analysis M2AN, Vol. 33 No 4, 1999, p. 651–672 Modélisation Mathématique et Analyse Numérique

DISCRETE APPROXIMATION OF THE MUMFORD-SHAH FUNCTIONAL IN DIMENSION TWO

Antonin Chambolle1 and Gianni Dal Maso2

Abstract. The Mumford-Shah functional, introduced to study image segmentation problems, is ap- proximated in the sense of Γ-convergence by a sequence of integral functionals defined on piecewise affine functions.

Résumé. La fonctionnelle de Mumford et Shah, proposée pour l’étude du problème de la segmentation d’images, est approchée au sens de la Γ-convergence par une suite de fonctionnelles intégrales définies sur des fonctions affines par morceaux.

AMS Subject Classification. 26A45, 49J45, 49Q20, 68U10.

Received: May 11, 1998.

1. Introduction

An important role in the variational approach to image segmentation problems (see [21–23]) is played by the Mumford-Shah functional ∫

Ω

|∇u(x)|2 dx + H1(Su) + ∫

Ω

|u(x)− g(x)|2 dx , (1.1)

where Ω is a bounded open subset of R2 with Lipschitz boundary, u belongs to the space SBV (Ω) of special functions of bounded variation on Ω (see Sect. 2), ∇u is the approximate gradient of u, Su is the set of essential discontinuity points of u, H1 is the one–dimensional Hausdorff measure, and g is a given function in L∞(Ω).

The main difficulty in the study of this functional is the presence of the length term H1(Su), and many works have been devoted to the problem of approximating this functional by simpler functionals defined on Sobolev spaces.

As pointed out in [10], it is impossible to obtain a variational approximation of (1.1), leading to the conver- gence of minimum points, by means of local integral functionals of the form∫

Ω

fε(∇u(x)) dx + ∫

Ω

|u(x)− g(x)|2 dx , (1.2)

Keywords and phrases. Free discontinuity problems, special bounded variation (SBV ) functions, Γ-convergence, finite elements.

1 CEREMADE, CNRS UMR 7534, Université de Paris-Dauphine, 75775 Paris Cedex 16, France. e-mail: chambolle@ceremade.dauphine.fr 2 SISSA, via Beirut 4, 34013 Trieste, Italy. e-mail: dalmaso@sissa.it

c© EDP Sciences, SMAI 1999

652 A. CHAMBOLLE AND G. DAL MASO

defined in the Sobolev space H1(Ω). Indeed, if such an approximation existed, (1.1) would be also the variational limit of the sequence of convex functionals∫

Ω

f∗∗ε (∇u(x)) dx + ∫

Ω

|u(x) − g(x)|2 dx ,

where f∗∗ε is the convex envelope of fε (see, e.g., Prop. 6.1 and Ex. 3.11 in [13]), and would therefore also be convex. This difficulty can be overcome by introducing an auxiliary variable as in the work of Ambrosio and Tortorelli (see [4,5]), while Braides and Dal Maso [10] recently found an approximation of (1.1) using non-local integral functionals.

In this work we consider only the two–dimensional case and present a different solution. We look for ap- proximations of the form (1.2), putting however a restriction on the function space on which it is defined. Our functionals are discrete, in the sense that they are defined on continuous piecewise affine functions that depend, for a fixed ε > 0, only on a finite number of parameters. On the other hand, the approach we propose is different from the finite-elements approximation introduced by Bellettini and Coscia [7], that is based on Ambrosio and Tortorelli’s functionals. It is also quite different from the finite-differences methods studied in [11], although some similarities exist and will appear clearly in some parts of the proofs.

The new approximation of the Mumford-Shah functional that we propose, based on adaptive finite elements, is studied in this paper from a theoretical point of view. In a forthcoming paper, Bourdin and Chambolle [9] address the problem of its numerical implementation, and show how it can be performed. Their algorithm alternates nonconvex optimization and mesh optimization. The way the optimal mesh is estimated relies a lot on the construction described in Section 4 and in the Appendix, that shows how the triangles need to be oriented in order to approximate well the energy of a given function.

In order to be more precise we need to introduce a few notations and definitions. A triangulation of Ω is, as usual, a finite family of closed triangles whose union contains Ω and such that, given any two triangles of this family, their intersection, if not empty, is exactly a vertex or an edge common to both triangles. Given some angle θ0, with 0 < θ0 ≤ 60◦, and a function ω(ε), with ω(ε) ≥ 6ε for any ε > 0 and limε→0+ ω(ε) = 0, we define, for any ε > 0, Tε(Ω) = Tε(Ω, ω, θ0) as the set of all triangulations of Ω made of triangles whose edges have length between ε and ω(ε), and whose angles are all greater than or equal to θ0. Then we consider the set Vε(Ω) of all continuous functions u : Ω → R for which there exists T ∈ Tε(Ω) such that u is affine on any triangle T ∈ T. We fix, like in [10], a non-decreasing continuous function f : [0,+∞)→ [0,+∞) such that

lim t→0+

f(t) t

= 1 and lim t→+∞

f(t) = f∞ < +∞ (1.3)

(the simplest case is the function f(t) = min{t, f∞}), and we set µ = f∞ sin θ0. Eventually, we fix some p ∈ [1,+∞) and define for every ε > 0 the functionals Fε : Lp(Ω)→ [0,+∞] and F : Lp(Ω)→ [0,+∞] by

Fε(u) =

1 ε

∫ Ω

f(ε|∇u(x)|2) dx , if u ∈ Vε(Ω),

+∞ , if u ∈ Lp(Ω) \ Vε(Ω), (1.4)

F (u) =

∫

Ω

|∇u(x)|2 dx + µH1(Su) , if u ∈ Lp(Ω) ∩GSBV (Ω),

+∞ , if u ∈ Lp(Ω) \GSBV (Ω), (1.5)

where GSBV (Ω) is the space of generalized special functions of bounded variation in Ω and Su is the jump set of u (we refer to the next section for the precise definitions). Note that it is not too difficult to check that, for any ε > 0, Vε(Ω) is weakly closed and that the functional Fε is weakly lower semi-continuous in Lp(Ω).

The main result of the paper is the proof of the Γ-convergence, as ε→ 0, of the functionals Fε to F .

DISCRETE APPROXIMATION OF THE MUMFORD-SHAH FUNCTIONAL IN DIMENSION TWO 653

Theorem 1.1. There exists a constant Θ, with 0 < Θ ≤ 60◦, such that, if 0 < θ0 < Θ, then the family (Fε) Γ-converges to F in Lp(Ω) as ε→ 0.

As a consequence of this theorem we will then establish the following result.

Theorem 1.2. Let g ∈ L∞(Ω), 0 < θ0 < Θ, and p > 1. For each ε > 0, let uε ∈ Vε(Ω) be a minimizer of

Fε(u) + ∫

Ω

|u(x)− g(x)|p dx .

Then, the family (uε)ε>0 is relatively compact in Lp(Ω) and any limit point u ∈ Lp(Ω) is a minimizer of

F (u) + ∫

Ω

|u(x)− g(x)|p dx .

Remark 1.3. The optimal value for Θ is not known, but the construction in the Appendix will show that Θ ≥ 45◦ − arctg (1/2) > 18◦. Remark 1.4. The construction in Section 4 will show that the conclusion of Theorem 1.2 remains valid (includ- ing the case p = 1) when the minimizers uε are computed under the additional constraint ‖uε‖L∞(Ω) ≤ ‖g‖L∞(Ω). Remark 1.5. The proofs in this paper may be easily adapted to the following variant. For any triangle T let hT be the smallest height of the triangle T . If u ∈ Vε(Ω) we say that T ∈ Tε(Ω) is adapted to u if u|T is affine for any T ∈ T; the set of the triangulations of Tε(Ω) adapted to u is denoted by Tε(u). If T ∈ Tε(u), let ∇uT be the gradient of u on any T ∈ T. For any u ∈ Lp(Ω) we define

Gε(u) =

inf

T∈Tε(u)

∑ T∈T |T ∩ Ω| 1

hT f(hT |∇uT |2) , if u ∈ Vε(Ω),

+∞ , otherwise, (1.6)

G(u) =

∫

Ω

|∇u(x)|2 dx + f∞H1(Su) , if u ∈ Lp(Ω) ∩GSBV (Ω),

+∞ , if u ∈ Lp(Ω) \GSBV (Ω). (1.7)

Then, Theorems 1.1 and 1.2 hold with Gε and G instead of Fε and F, provided ω(ε) ≤ c ε for some c > 0 (another possibility is to replace in the above formulas hT with min{hT , c ε}, in which case this constraint on ω(ε) is no more needed). In this paper we prefer to use Fε, whose meaning is more immediate, but the numerical implementation (see [9]) of Gε is much simpler, as a better value of the minimum can be reached even when the triangulation is not perfectly optimized along the discontinuity sets. Whereas, with the functional Fε, the optimal triangulation must be made of the smallest possible triangles aligned along the discontinuity, as shown in Section 4.

In the next section we will define precisely the spaces SBV (Ω) and GSBV (Ω), and give the definition of the Γ-convergence. Then the remaining three sections will be devoted to the proof of Theorems 1.1 and 1.2.

2. Notation and preliminaries

The scalar product of x, y ∈ R2 is denoted by 〈x, y〉 and the Euclidean norm by |x|. The open ball with centre x and radius ρ is indicated by Bρ(x). Given some open set A ⊂ R2 and ρ > 0, Aρ is the set {x ∈ A : dist(x, ∂A) > ρ}, where dist(x, ∂A) is the distance from x to ∂A. The Lebesgue measure and the one–dimensional Hausdorff measure of a Borel set B ⊂ R2 are denoted by |B| and H1(B), respectively. For the general properties of the Hausdorff measure we refer to [17,18]. We use standard notation for Lebesgue spaces Lp(Ω) and Sobolev spaces H1(Ω).

654

Recommended

*View more*