Efficient Global Minimization for the Multiphase

Chan-Vese Model of Image Segmentation ∗

Egil Bae† Xue-Cheng Tai‡

Abstract
The Mumford-Shah model is an important variational image segmen-
tation model. A popular multiphase level set approach, the Chan-Vese
model, was developed as a numerical realization by representing the phases
by several overlapping level set functions. Recently, a variant representa-
tion of the Chan-Vese model with binary level set functions was proposed.
In both approaches, the gradient descent equations had to be solved nu-
merically, a procedure which is slow and has the potential of getting stuck
in a local minima.
In this work, we develop an efficient and global minimization method
for a discrete version of the level set representation of the Chan-Vese model
with 4 regions (phases), based on graph cuts. If the average intensity val-
ues of the different regions are sufficiently evenly distributed, the energy
function is submodular. It is shown theoretically and experimentally that
the condition is expected to hold for the most commonly used data terms.
We have also developed a method for minimizing nonsubmodular func-
tions, that can produce global solutions in practice should the condition
not be satisfied, which may happen for the L1 data term.

1 Introduction
Multiphase image segmentation is a fundamental problem in image processing.
Variational models such as the Mumford-Shah model [27] are powerful for this
task, but efficient numerical computation of a global minimum is a big challenge.
The level set method [14, 29] is a powerful tool which can used for numerical
realization. It was first proposed for the Mumford-Shah model in [10] for two
phases and [33] for multiple phases. Both approaches have the disadvantage of
slow convergence and potential of getting stuck in a local minima.
∗ Support from the Norwegian Research Council (eVita project 166075), National Science

Foundation of Singapore (NRF2007IDM-IDM002-010) and Ministry of Education of Singapore

(Moe Tier 2 T207B2202) are gratefully acknowledged.
† Department of Mathematics, University of Bergen, Norway ([email protected]).
‡ Department of Mathematics, University of Bergen, Norway and Division of Mathematical

Sciences, School of Physical and Mathematical Sciences, Nanyang Technological University,

Singapore. ([email protected]).

1
 Graph cuts from combinatorial optimization [15, 7, 16, 4, 20, 21, 3] is another
technique which can perform image segmentation by minimizing certain discrete
energy functions. In the recent years, the relationship between graph cuts and
continuous variational problems have been much explored [5, 6, 12, 13]. It turns
out graph cuts are very similar to the level set method, and can be used for
many variational problems with the advantage of a much higher efficiency and
ability to find global minima. It can be applied to the 2-phase Mumford-Shah
model [11, 35], but for multiple phases it is probably not possible to find the
exact, global minimum in polynomial time as this is an NP-hard problem. The
usual approach to minimization problems with several regions is some heuristic
method for finding an approximate, local minimum. Most popular in computer
vision are the alpha-expansion and alpha-beta swap algorithms [7]. Recently,
also convex formulations of the continuous multiphase problem have been made
in [30, 23, 34] by relaxing the integrality constraint. A suboptimal solution is
found by converting the real valued relaxed solution to an integral one (e.g. by
thresholding).
In this paper we propose a method for globally and efficiently minimizing
the energy in the Mumford-Shah model in the multiphase level set framework
of Vese and Chan [33] by using binary level set functions as in [24]. Since the
length term is slightly approximated in this framework, global minimization is
not NP hard.
We will construct a graph such that the discrete variational problem can be
minimized exactly by finding a minimum cut on the graph, provided the energy
function is submodular. A sufficient condition on the data term is derived for
when the energy is submodular. It is shown theoretically and experimentally
that the condition is expected to hold for the most commonly used data term.
The submodularity of the energy function depends on how evenly the average
intensity values of each region are distributed and may sometimes be violated
under an L1 data fitting term. To handle these cases, we have developed a
method for minimizing non-submodular functions with particular emphasis on
this energy function. It is shown in experiments that the algorithm can effi-
ciently compute global solutions in practice, but we cannot prove it will always
do so (which would conflict with NP-hardness of the problem).
The global optimization framework applies if the average intensity values of
each region are fixed. One can also simultaneously minimize with respect to
the average intensities values, by an alternating algorithm as in [2]. Although a
global minimum with respect to the average intensities is not guaranteed, such
an algorithm is much more robust to initialization.
Note that in contrast to alpha-expansion and alpha-beta swap, the approx-
imation is done in the model rather than in the minimization method. Our
approach only requires to find a minimum cut on a graph once, while alpha
expansion and alpha-beta swap need to iteratively solve a sequence of minimum
cut problems. By analyzing the complexity it can be easily seen that our method
is more efficient.
In this work we will focus fully on the case of 4 or less phases, but aim to
generalize to more phases later. Nevertheless, these are important cases since

2
by the four colour theorem, four phases in theory suffices to segment any 2D
image. This can potentially be exploited in an algorithm in the future, by
assigning different constant values to each disconnect component of the phases.
The paper is organized as follows: Section 1.1 reviews the Mumford-Shah
model, the Chan-Vese model and the different level set representations. Section
2 presents the new global minimization approach for the multiphase Chan-Vese
model. In Section 3 some theoretical analysis of the submodularity condition is
presented. Section 4 presents algorithms for minimizing non-submodular energy
functions.
Section 5 presents a local minimization approach for determining the average
intensity values of the phases, while numerical experiments are presented in
Section 6.

1.1 The Mumford-Shah model and its level set represen-

tation
Image segmentation is the task of partitioning the image domain Ω into a set
of n meaningful disjoint regions {Ωi }ni=1 .The Mumford-Shah model [27] is an
established image segmentation model with a wide range of applications. An
energy functional to be minimized is defined over the regions {Ωi }ni=1 , and an
approximation image u of the input image u0 . In an especially popular form, u
is assumed to be constant within each region Ωi , in which case the model reads
n 
 n
 
0 β
min EMS ({ci }, {Ωi }) = |ci − u | dx + ν ds, (1)
{ci },{Ωi } Ωi ∂Ωi
i=1 i=1

where ∂Ωi is the boundary of Ωi . The power β is usually chosen as β = 2.

As a numerical realization, Chan and Vese [10, 33] proposed to represent the
above functional with level set functions, and solve the resulting gradient descent
equations numerically. For two phases (n = 2) the level set representation
yielded the variational problem

min ν |∇H(φ)| + {H(φ)|c1 − u0 |β + (1 − H(φ))|c2 − u0 |β }dx, (2)
φ,c1 ,c2 Ω

where H(·) : R → R is the Heaviside function H(x) = 0 if x < 0 and H(x) = 1 if

x ≥ 0. The multiphase case was handled by introducing more level set functions.
By using m = log2 (n) level set functions, denoted φ1 , ..., φm , n phases could be
represented. An important special case is the representation of 4 phases by two
level set functions φ1 ,φ2 , as in Table 1. The energy functional could then be
written
 
1 2 1
1
min
2
E CV (φ , φ , c 1 , ..., c 4 ) = ν |∇H(φ )| + ν |∇H(φ2 )| (3)
φ ,φ ,c1 ,...,c4 Ω Ω

+ {H(φ1 )H(φ2 )|c2 − u0 |β + H(φ1 )(1 − H(φ2 ))|c1 − u0 |β )
Ω

3
 Traditional level set functions Binary level set functions
x ∈ phase 1 iff φ1 (x) > 0, φ2 (x) < 0 φ1 (x) = 1, φ2 (x) = 0
x ∈ phase 2 iff φ1 (x) > 0, φ2 (x) > 0 φ1 (x) = 1, φ2 (x) = 1
x ∈ phase 3 iff φ1 (x) < 0, φ2 (x) < 0 φ1 (x) = 0, φ2 (x) = 0
x ∈ phase 4 iff φ1 (x) < 0, φ2 (x) > 0 φ1 (x) = 0, φ2 (x) = 1

Table 1: Representation of four phases by traditional and binary level set func-
tions (note: a little permutation compared to the original paper [33]).

+(1 − H(φ1 ))H(φ2 )|c4 − u0 |β dx + (1 − H(φ1 ))(1 − H(φ2 ))|c3 − u0 |β }dx.

Note also that we have made a permutation in the interpretation of the phases
compared to [33]. The energy is still exactly the same for all possible solutions.
This permutation is crucial for making the corresponding discrete energy func-
tion submodular. Note also that the length term in (1) is slightly approximated,
since some of the boundaries are counted twice. In fact there are a total of 6
types of boundaries between two different regions. In this model only two of
them are counted twice, while the remaining 4 are counted once. This is very
close to the ideal situation where each boundary is counted once.
The functional in the variational problem (3) is highly non-convex, even for
fixed constant values c1 , ..., c4 . The traditional minimization approach of solving
the gradient descent equations can therefore easily get stuck in a local minima.
Furthermore, the numerical solution of the gradient descent PDEs is expensive
computationally.
In [26, 24], the same multiphase model was formulated using binary level set
functions φ1 , φ2 ∈ D = {φ | φ : Ω → {0, 1}}, representing the phases as in Table
1. This resulted in the energy functional
 
1 2 1
1 2
min ECV (φ , φ , c 1 , ..., c 4 ) = ν |∇φ |dx+ν |∇φ2 |dx+E data (φ1 , φ2 ),
φ ,φ ∈D,c1 ,...,c4 Ω Ω
(4)
where

E data (φ1 , φ2 ) = {φ1 φ2 |c2 − u0 |β + φ1 (1 − φ2 )|c1 − u0 |β )
Ω

+(1 − φ1 )φ2 |c4 − u0 |β + (1 − φ1 )(1 − φ2 )|c3 − u0 |β }dx.

The functional was written in a slightly different form in [26, 24], but is exactly
equal to the above in case φ1 and φ2 are binary functions. The constraint D
was represented by the polynomials K(φ1 ) = 0 and K(φ2 ) = 0, where K(φ) =
φ(1 − φ). Minimization of this constrained problem was carried out by the
augmented lagrangian method. Since both the side constraints were non-convex,
global minimization could not be guaranteed. Also, convergence was slow just
as in the traditional level set approach. A similar approach could also be used
for finding a local minimum with exact curve lengths [25].

4
 Let us mention that a method often referred to as continuous graph cut can
be used to globally minimize the Mumford Shah model in case of two phases.
By letting φ ∈ D, this model can be written

min ν |∇φ|dx + {φ|c1 − u0 |β + (1 − φ)|c2 − u0 |β }dx. (5)
φ∈D,c1 ,c2 Ω

The idea, presented in [28] is to relax the constraint D by the convex constraint
D = {φ | φ : Ω → [0, 1]}. It was shown that thresholding this solution at almost
any threshold in (0, 1] yields the optimal solution within D. Since (5) is convex,
this procedure would yield the globally optimal solution.
One might immediately think the same idea could be extended to the multi-
phase case by iteratively minimizing (4) for φ1 and φ2 in D and finally threshold
the results. However, since E data (φ1 , φ2 ) is not convex with respect to φ1 and
φ2 , the minimization would not be global.
In general, discrete graph cut has the disadvantage of some metrication ar-
tifacts over continuous graph cuts. However, discrete graph cuts is faster and
can elegantly be used for minimization problems with non-local operators. The
method we propose can easily be generalized to minimize non-local measure-
ments of the curve lengths as was done for two phases in [8].
In the next section we will propose a method which globally minimizes (4) for
fixed constant values c1 , ..., c4 . This new approach is also shown to be very su-
perior in terms of efficiency compared to gradient descent. We start by deriving
a discretizatation of (4).

1.2 Discretization of energy functional

Instead of discretizing the Euler-Lagrange equations, we will discretize the vari-
ational problem (4). In the next section we show how to minimize the resulting
discrete energy function exactly by graph cuts. Let us first mention there are
two variants of the total variation term. The isotropic variant, by using 2-norm
  
T V2 (φ) = |∇φ|2 dx = |φx1 |2 + |φx2 |2 dx (6)
Ω Ω

and the anisotropic variant, by using 1-norm

 
T V1 (φ) = |∇φ|1 dx = |φx1 | + |φx2 | dx. (7)
Ω Ω

Only the anisotropic variant is graph representable and will be considered here.
A more isotropic graph representable version can be obtained by splitting T V1
using the original gradient operator, and one rotated counterclockwise π/4 ra-
dians 
1  
T V1, 4 (φ) =
π |∇φ(x)|1 + |R π4 ∇φ(x)|1 dx, (8)
2 Ω
where R π4 ∇ is the gradient in the rotated coordinate system. It is also possible
to create even more isotropic versions by considering more such rotations.

5
 Let P = {(i, j) ⊂ Z2 } denote the set of grid points. For each p = (i, j) ∈ P,
the neighborhood system Npk ⊂ P is defined as

Np4 = {(i ± 1, j), (i, j ± 1)} ∩ P

Np8 = {(i ± 1, j), (i, j ± 1), (i ± 1, j ± 1)} ∩ P.

The discrete energy function can be written
   
1 2 1 1
1 2
min E d (φ , φ , c 1 , ..., c 4 ) = ν wpq |φp −φq |+ν wpq |φ2p −φ2q |
φ ,φ ∈D,c1 ,...,c4
p∈P q∈Npk p∈P q∈Npk

 (9)
+ Epdata (φ1p , φ2p ),
p∈P

where
Epdata (φ1p , φ2p ) = {φ1p φ2p |c2 − u0p |β + φ1p (1 − φ2p )|c1 − u0p |β )
+(1 − φ1p )φ2p |c4 − u0p |β + (1 − φ1p )(1 − φ2p )|c3 − u0p |β },
and k = 4 for T V1 and k = 8 for T V1, π4 . The weights wpq are then given by wpq =
4δ 2
k||p−q||2 .
Similar weights can also be derived from the Cauchy-Crofton formula
of integral geometry as was done for two phases in [5]. It can furthermore be
proved that as the mesh size decreases and the size of the neighborhood system
increases, the minimizers of the discrete energy function converges to minimizers
of the continuous energy functional.

2 Graph cut minimization

We will show that the discrete problem (9) can be minimized globally by finding
the minimum cut on a specially designed graph. This is possible provided the
constant values c1 , ..., c4 are sufficiently evenly distributed. We show that such
a distribution makes the discrete energy function sub-modular. Some analysis
of the condition is given in Section 3, where it is argued it will be satisfied in
practice for the most commonly used data terms. In Section 4, an algorithm is
developed for minimizing non-submodular functions with particular emphasize
on functions of the form (4).

2.1 Brief overview of graph cuts in computer vision

Graph cut is a well known optimization problem. Due to a duality theorem by
Ford and Fulkerson [22], there are several fast algorithms for this problem. It
was introduced as a computer vision tool by Greig et. al. [16] in connection
with markov random fields [15] and has later been studied by Kolmogorov et. al.
[4, 20]. Its applications range from stereo vision [19], segmentation [3, 17, 35, 11]
to noise removal [12, 13, 9].

6
 A graph G = (V, E) is a set of vertices V and a set of edges E. We let (a, b)
denote the directed edge going from vertex a to vertex b, and let c(a, b) denote
the capacity/cost/weight on this edge. In the graph cut scenario there are two
distinguished vertices in V, called the source {s} and the sink {t}. A cut on G
is a partitioning of the vertices V into two disjoint connected sets (Vs , Vt ) such
that s ∈ Vs and t ∈ Vt . The cost of the cut is defined as

c(Vs , Vt ) = c(i, j).
(i,j)∈E s.t. i∈Vs ,j∈Vt

A flow f on G is a function f : E → R. For a given flow, the residual

capacities are defined as R(e) = c(e) − f (e) ∀e ∈ E. The max flow problem is to
find maximum amount of flow that can be pushed from {s} to {t}, under flow
conservation constraint at each vertex. The theorem of Ford and Fulkerson says
this is the dual to the problem of finding the cut of minimum cost on G, the min-
cut problem. Therefore, efficient algorithms for finding max-flow, such as the
augmented paths method [22] can be used for finding minimum cuts in graphs.
An efficient implementation of this algorithm specialized for image processing
problems can be found in [4]. This algorithm, which is available on-line has
been used in our experiments.
In computer vision this has been exploited for minimizing energy functions
of the form  
min m E(x) = E i (xi ) + E i,j (xi , xj ).
x∈{0,1}
i i<j

Typically, i = 1, ..., m denotes the set of grid points and x contains one binary
variable for each grid point. In order to be representable as a cut on a graph,
it is required that the energy function is submodular (or regular) [20, 15], i.e.

E i,j (0, 0) + E i,j (1, 1) ≤ E i,j (0, 1) + E i,j (1, 0), ∀i < j

2.2 Graph construction

Observe that the energy function E is composed of pairwise interaction terms
between binary variables. Such energy functions can be minimized exactly via
graph cuts, provided the pairwise interaction potentials are submodular [20, 15].
In particular, this requires the data term to satisfy

Epdata (1, 1) + Epdata (0, 0) ≤ Epdata (1, 0) + Epdata (0, 1) (10)

We will construct a graph G such that there is a one-to-one correspondence

between cuts on G and the level set functions φ1 and φ2 . Furthermore, the
minimum cost cut will correspond to the level set functions φ1 and φ2 minimizing
the energy (9).

1 2
min c(Vs , Vt ) = min
1 2
Ed (φ , φ , c 1 , ..., c 4 ) + σp . (11)
(Vs ,Vt ) φ ,φ
p∈P

7
 (a) (b)

Figure 1: (a) The graph corresponding to the data term at one grid point p.
(b) A sketch of the graph corresponding to the energy function of a 1D signal
of two grid points p and q, red: data edges, blue: regularization edges.

where σp ∈ R are fixed for each p ∈ P. In the graph, two vertices are associated
to each grid point p ∈ P. They are denoted vp,1 and vp,2 , and correspond to
each of the level set functions φ1 and φ2 . Hence the set of vertices is formally
defined as
V = {vp,i | p ∈ P, i = 1, 2} ∪ {s} ∪ {t}. (12)
The edges are constructed such that the relationship (11) is satisfied. We begin
with the edges constituting the data term of (9). For each grid point p ∈ P they
are defined as

ED (p) = (s, vp,1 ) ∪ (s, vp,2 ) ∪ (vp,1 , t) ∪ (vp,2 , t) ∪ (vp,1 , vp,2 ) ∪ (vp,2 , vp,1 ). (13)

The set of all data edges are denoted ED and defined as ∪p∈P ED (p). The edges
corresponding to the regularization term are defined as

ER = {(vp,1 , vq,1 ), (vp,2 , vq,2 ) ∀p, q ⊂ P s.t.q ∈ Npk }. (14)

For any cut (Vs , Vt ), the corresponding level set functions are defined by
 
1 1 if vp,1 ∈ Vs , 2 1 if vp,2 ∈ Vs ,
φp = φp = (15)
0 if vp,1 ∈ Vt , 0 if vp,2 ∈ Vt .

Weights are assigned to the edges such that the relationship (11) is satisfied.
Weights on the regularization edges are simply given by

∀p ∈ P, q ∈ Npk .
c (vp,1 , vq,1 ) = c (vq,1 , vp,1 ) = c (vp,2 , vq,2 ) = c (vq,2 , vp,2 ) = νwpq ,
(16)
We now concentrate on the weights on data edges ED . For grid point p ∈ P, let

A(p) = c(vp,1 , t), B(p) = c(vp,2 , t), C(p) = c(s, vp,1 ),

D(p) = c(s, vp,2 ), E(p) = c(vp,1 , vp,2 ), F (p) = c(vp,2 , vp,1 ).

8
In Figure 1(a) the graph corresponding to an image of one pixel p is shown. It
is clear that these weights must satisfy
⎧
⎪
⎪ A(p) + B(p) = |c2 − u0p |β + σp
⎨
C(p) + D(p) = |c3 − u0p |β + σp
(17)
⎪
⎪ A(p) + E(p) + D(p) = |c1 − u0p |β + σp
⎩ 0 β
B(p) + F (p) + C(p) = |c4 − up | + σp

This is a non-singular linear system for the weights A(p), B(p), C(p), D(p), E(p), F (p).
Negative weights are not allowed. By choosing σp large enough there will
exist a solution with A(p), B(p), C(p), D(p) ≥ 0. However, the requirement
E(p), F (p) ≥ 0 implies that

|c1 − u0p |β + |c4 − u0p |β = A(p) + B(p) + C(p) + D(p) + E(p) + F (p) − 2σp

≥ A(p) + B(p) + C(p) + D(p) − 2σp = |c2 − u0p |β + |c3 − u0p |β .

This condition must hold for all grid points p ∈ P, which is exactly the sub-
modular condition (10). Hence, the following condition on the constant values
c1 , ..., c4 must be satisfied

|c2 − I|β + |c3 − I|β ≤ |c1 − I|β + |c4 − I|β , ∀ I ∈ [0, L], (18)

where L is the maximum intensity value.

A detailed analysis of the condition (18) is given in Section 3.
Assuming (18) holds, the linear system (17) has infinitely many solutions. It
was shown in [20] that at most three edges are required for representing a general
submodular term of two binary variables. Therefore, it is possible to pick a
solution such that at least three of the weights A(p), B(p), C(p), D(p), E(p), F (p)
in ED (p) become zero for each p ∈ P. The construction of the solution is as
follows
A(p) = max{|c2 − u0p |β − |c4 − u0p |β , 0}, (19)
C(p) = max{|c4 − u0p |β − |u0p − c2 |β , 0} (20)
B(p) = max{|c4 − u0p |β − |c3 − u0p |β , 0}, (21)
D(p) = max{|c3 − u0p |β − |c4 − u0p |β , 0}, (22)
E(p) = |c1 − u0p |β + |c4 − u0p |β − |c2 − u0p |β − |c3 − u0p |β , F (p) = 0. (23)
The value σp is given implicitly by this solution.
Therefore, by analyzing the complexity of the method in the augmented
paths framework, it is easily seen that the computational cost is equal to the
cost of one single iteration of the alpha expansion method.
Note finally that three phase segmentation is a special case that can be
handled by putting infinite cost to one of the four possible solutions, i.e. E(p) =
∞ or F (p) = ∞.

9
 (a) (b)

(c) (d)

Figure 2: (a), (b) and (c) distributions of c which makes energy function sub-
modular for all β. (d) distribution of c which may make energy function non-
submodular for small β

3 Analysis of submodular condition

The condition (18) says something about how evenly {ci }4i=1 are distributed.
First we characterize situations for which (10) is always satisfied.
Proposition 1. Let 0 ≤ c1 < c2 < c3 < c4 . (18) is satisfied for all I ∈
[ c2 −c
2 ,
1 c4 −c3
2 ] for any β ≥ 1.

Proposition 2. Let 0 ≤ c1 < c2 < c3 < c4 . (18) is satisfied for any β ≥ 1 if

c2 − c1 = c4 − c3 .
Further, it can be observed that (18) becomes less strict as β increases, as
the next two results show.
Proposition 3. Let 0 ≤ c1 < c2 < c3 < c4 . If (18) is satisfied for some β ≥ 1,
then (10) is satisfied for all powers α ≥ β.
Proposition 4. Let 0 ≤ c1 < c2 < c3 < c4 . There exists a B ∈ N such that
(18) is satisfied for any β ≥ B.
In fact we have observed that for β = 2, (18) is always satisfied in practice
for optimal constant values.
Figure 2 shows examples where the condition is satisfied and may fail. Prop.
3 is illustrated in Figure 2 (b) and (c). Figure 2 (d) shows the only possibility in
which (18) may be violated, i.e. c1 , c2 , c3 are clustered compared to c4 (the sym-
metrical version would also be a problem). However, the model (4) will disfavor
solutions where the constants are clustered. Experiments will demonstrate that
under L2 data fidelity (18) is always satisfied for optimal values of c. Under L1
data fidelity, it may be more easily violated.
In case (18) does not hold at some p ∈ P, the energy function is non-
submodular. Not only does this mean it cannot in general be minimized by
graph cut. It also implies the minimization problem is NP-hard, hence no
algorithm exist that can solve the problem in polynomial time (unless P = N P ).

10
4 Minimization of non-submodular energy func-
tions
Assume that the submodularity condition (18) is not satisfied. We will develop
a method for minimizing non-submodular energy functions with particular em-
phasize on energy functions of the form (9). The algorithm cannot be guaran-
teed to always find a global solution, but works well in practice. Minimization of
non-submodular functions via graph cuts has been investigated previously, see
[18] for a review. The usual idea is to develop a method for determining most
of the variables, while leaving some of the variables undetermined. We instead
aim to determine all the variables. Even when (18) does not hold, the energy
function is ”almost submodular”, which we believe explains why the following
very efficient algorithms work so well in practice.
Assume that

|c2 − u0p |β + |c3 − u0p |β > |c1 − u0p |β + |c4 − u0p |β ,

for some p ∈ P. In this case the linear system (17) has a solution only if either
E(p) < 0 or F (p) < 0, in which case one of the edges, (vp,1 , vp,2 ) or (vp,2 , vp,1 ),
will have negative weight. In order to construct the solution we consider two
cases. If u0p > c3 , then

E(p) = |c1 − u0p |β + |c4 − u0p |β − |c2 − u0p |β − |c3 − u0p |β , F (p) = 0 (24)

A(p) = max{|c2 − u0p |β − |c4 − u0p |β , 0} − E(p), (25)

C(p) = max{|c4 − u0p |β − |u0p − c2 |β , 0} − E(p), (26)
B(p) = max{|c4 − u0p |β − |c3 − u0p |β , 0} − E(p), (27)
D(p) = max{|c3 − u0p |β − |c4 − u0p |β , 0} − E(p), (28)
in which case E(p) < 0. If u0p < c2 , then

F (p) = |c1 − u0p |β + |c4 − u0p |β − |c2 − u0p |β − |c3 − u0p |β , E(p) = 0 (29)

A(p) = max{|c1 − u0p |β − |c3 − u0p |β , 0} − F (p), (30)

C(p) = max{|c3 − u0p |β − |u0p − c1 |β , 0} − F (p), (31)
B(p) = max{|c2 − u0p |β − |c1 − u0p |β , 0} − F (p), (32)
D(p) = max{|c1 − u0p |β − |c2 − u0p |β , 0} − F (p), (33)
in which case F (p) < 0. By Prop 1, the condition holds whenever u0p ∈ [c2 , c3 ].
In this section we let G denote the graph with data edges set according to
(24)-(33) (some of which may be negative) and regularization edges set to (16).

11
4.1 Truncation of non-submodular terms
It is difficult to interpret what is physically meant by max flow on a graph with
negative edge weights. The concept of min-cut, on the other hand, certainly
have a meaning even if some of the edges have negative weight. If all the
edges have negative weight, the min-cut problem becomes equivalent to the
max-cut problem on a graph with negated edge weights. The first step of the
method finds a good feasible solution, and therefore also a good upper bound
on the objective function (9). It seems that most often this feasible solution is
in fact the optimal solution. All edges with negative weight will be removed,
resulting in a new graph G. It has been observed in [31] that removing negative
edges, often called truncation, can be effective in minimizing non-submodular
functions. We will see that this applies especially well to our energy function.
Furthermore, we will derive a criterion for when the minimum cut on the graph
with removed edges of negative weight is also a minimum cut on the original
graph with negative edge weights.
Let G be the graph identically to G except that all edges of negative weight
are removed. The minimum cut on G can be easily computed by max-flow. As
discussed in the previous section, the condition (18) may only be violated if
c1 , c2 , c3 are close to each other compared to c4 and u0p at p ∈ P is close to c4 .
Measured by the data term, the worst assignment of p is to phase 1, which has
the cost |c1 − u0p |β . By removing the edge with negative weight E(p) < 0, the
cost of this assignment becomes even higher |c1 − u0p |β − E(p). Alternatively,
if c2 , c3 , c4 are close to each other compared to c1 and u0p is close to c1 then
F (p) < 0. By removing the edge with negative weight, the cost of the worst
assignment of u0p becomes higher |c4 −u0p |β −F (p). We therefore expect minimum
cuts on G to be almost identical to minimum cuts on G. Define the sets

P 1 = {p ∈ P | E(p) < 0, F (p) ≥ 0},

P 2 = {p ∈ P | F (p) < 0, E(p) ≥ 0},

consisting of all p ∈ P for which either E(p) < 0 or F (p) < 0.
Assume the maximum flow has been computed on G, let RA (p), RB (p), RC (p), RD (p)
denote the residual capacities on the edges (vp,1 , t), (vp,2 , t), (s, vp,1 ), (s, vp,2 ) re-
spectively. The following theorem gives a criterion for when the minimum cut
on G yields the optimal solution of the original problem.
Theorem 5. Let G be a graph as defined in (12)-(14) and (16), with weights
A(p), B(p), C(p), D(p), E(p), F (p) satisfying (17). Let G be the graph with
weights as in G, with the exception c(vp,1 , vp,2 ) = 0 ∀p ∈ P 1 and c(vp,2 , vp,1 ) = 0
∀p ∈ P 2 .
Assume the maximum flow has been computed on the graph G. If

RA (p) + RD (p) ≥ −E(p), ∀p ∈ P 1 and RB (p) + RC (p) ≥ −F (p), ∀p ∈ P 2 ,

(34)
then min-cut (G) = min-cut (G).

12
 (a) G (b) G

Figure 3: Illustration of graph G in case E(p) < 0.

Proof. We will create a graph G of only positive edge weights, such that the
minimum cut problem on G is a relaxation of the minimum cut problem on G.
The graph G is constructed with weights as in G with the following exceptions

c(vp,1 , t) = A(p) − RA (p), ∀p ∈ P 1 ,

c(s, vp,2 ) = D(p) − RD (p), ∀p ∈ P 1

c(vp,2 , t) = B(p) − RB (p), ∀p ∈ P 2 ,
c(s, vp,1 ) = C(p) − RC (p), ∀p ∈ P 2 .
We will first show min-cut(G) ≤ min-cut(G) ≤ min-cut(G). The right inequality
follows because all the edges in the graph G have greater or equal weight than the
edges in the graph G. To prove the left inequality, observe that only data edges
for p ∈ P1 ∪ P2 differ between G and G. For each p ∈ P1 there are 4 possibilities
for the cut (Vs , Vt ). Since RA (p), RB (p), RC (p), RD (p) ≥ 0, the cost of all the
3 cuts vp,1 , vp,2 ∈ Vs , vp,1 , vp,2 ∈ Vt and vp,1 ∈ Vt , vp,2 ∈ Vs are lower in G than
in G. The last cut vp,1 ∈ Vs , vp,2 ∈ Vt has the cost A(p) + B(p) − E(p) in the
G and the cost A(p) + D(p) − (RA (p) + RD (p)) ≤ A(p) + D(p) + E(p) in the
graph G. The same argument shows that all possible cuts have a lower or equal
cost in G than in G for p ∈ P2 .
Both G and G have only positive edge weights. Since all the edges have
greater or equal weight in G than in G it follows that

max-flow(G) ≤ max-flow(G).

Hence, since the max flow on G is feasible on G it is also optimal on G. Therefore,

by duality min-cut(G) = min-cut(G) which implies min-cut(G) = min-cut(G).

13
 Therefore, by computing the max flow on G and examining the residual
capacities for criterion (34), we can check whether the solution is optimal on
G. Most often it is possible to stop at this stage, since the residual capacity
conditions are satisfied everywhere.

4.2 Refinement of truncated solution

If the regularization parameter is set extremely high, the residual criterion (34)
may be violated at a small set of the pixels. In that case one could accept
the computed solution as suboptimal. We have also developed an algorithm
which iterates upon the cut on G until it finds a cut on G. We cannot prove the
algorithm always converges (which would conflict with the NP-hardness of the
problem), but it works well in practice.
If the residual capacity conditions are violated, there is a possibility the
optimal assignment is φ1 (p) = 0, φ2 (p) = 1 if p ∈ P1 and φ1 (p) = 1, φ2 (p) = 0 if
p ∈ P2 , even if a different assignment was produced by the cut on the graph G.
In the first step of the algorithm, the weights on the data edges at p are modified
such that the cost of the above assignments are correct, at the expense of one
of the remaining 3 assignments. There are possibilities for several variants.
In the variant below, the weights are modified such that the cost of either
assignment φ1 (p), φ2 (p) = 0 or φ1 (p), φ2 (p) = 1 are reduced, while the remaining
assignments are correct. Which assignment to be modified, is determined by
examining the cut on the graph G. The cut on the modified graph is expected
to be closer to a cut on G. If the cut on the new graph may causes either: (1)
the residual capacity condition to be violated at some new p ∈ P1 ∪ P2 , (2)
the assignment of φ1 (p), φ2 (p) to a region of reduced cost, one cannot guarantee
the obtained cut is also optimal on G. In that case, the same procedure can be
reiterated until neither (1) or (2) are violated.
The algorithm creates a sequence of graphs {Gi }ni=1 , with G0 = G, of only
positive edge weights, such that the min-cut on Gi converges to a min-cut on
G. The graphs satisfy min-cut(G0 ) = min-cut(G), min-cut(Gi ) ≤ min-cut(G) for
all i, and min-cut(Gn ) = min-cut(G). For a given flow on Gi we define two new
sets Pi1 ⊆ P 1 and Pi2 ⊆ P 2

Pi1 = {p ∈ P 1 | RA
i i
(p) + RD (p) < −E(p)},

Pi2 = {p ∈ P 2 | RB
i i
(p) + RC (p) < −F (p)},
i i i i
where RA (p), RB (p), RC (p) and RD (p) are the residual capacities in graph Gi
on edges (vp,1 , t), (vp,2 , t), (s, vp,1 ), (s, vp,2 ) respectively.
By construction, a sufficient stopping criterion that ensures both (1) and (2)
above is to require Gi = Gi−1 , that is, the weights on all edges of Gi and Gi−1
are equal. The algorithm is as follows.

Algorithm 1:
G0 = G, G−1 = ∅, i = 0
Find max flow on G0 , update P01 and P02

14
 if (P01 and P02 are empty)
stop, optimal solution found
else:
while(Gi = Gi−1 ){
1. Construct Gi+1 as in G except for the following weights

for all p ∈ Pi1

if(vp,1 ∈ Vt and vp,2 ∈ Vt in Gi ): set c(vp,1 , t) = A(p) + E(p)
if(vp,1 ∈ Vs and vp,2 ∈ Vs in Gi ): set c(s, vp,2 ) = D(p) + E(p)
if(vp,1 ∈ Vs and vp,2 ∈ Vt in Gi ): set c(s, vp,1 ) = A(p) + E(p)
if(vp,1 ∈ Vt and vp,2 ∈ Vs in Gi ): set c(s, vp,1 ) = D(p) + E(p)

for all p ∈ Pi2

if(vp,1 ∈ Vt and vp,2 ∈ Vt in Gi ): set c(vp,2 , t) = B(p) + F (p)
if(vp,1 ∈ Vs and vp,2 ∈ Vs in Gi ): set c(s, vp,1 ) = C(p) + F (p)
if(vp,1 ∈ Vs and vp,2 ∈ Vt in Gi ): set c(s, vp,2 ) = B(p) + F (p)
if(vp,1 ∈ Vt and vp,2 ∈ Vs in Gi ): set c(s, vp,2 ) = C(p) + F (p)

2. Find max-flow on Gi+1

1 2
3. Update Pi+1 and Pi+1 by examining residual capacities in
Gi+1 4. i ← i + 1
}

Theorem 6. Let Gn be the graph at termination of Algorithm 1 and let (Vs , Vt )

be the minimum cut on Gn computed at the last iteration. Then (Vs , Vt ) is a
minimum cut on G.
Proof. If the algorithm terminates with G0 , optimality was proved in theorem 5.
Assume therefore n ≥ 1. The proof follows some of the same ideas as the proof
of theorem 5. We will use Gn to construct a graph G such that the minimum
cut problem on G is a relaxation of the minimum cut problem on G. Observe
first that since Gn = Gn−1 , the minimum cut on Gn is feasible, no edges in the
cut have a reduced cost. Therefore, min-cut(Gn ) ≥ min-cut(G)
The graph G is constructed with weights as in Gn except
n
c(vp,1 , t) = A(p) − RA (p), ∀p ∈ P 1 \Pn1 ,
n
c(s, vp,2 ) = D(p) − RD (p), ∀p ∈ P 1 \Pn1
n
c(vp,2 , t) = B(p) − RB (p), ∀p ∈ P 2 \Pn2 ,
n
c(s, vp,1 ) = C(p) − RC (p), ∀p ∈ P 2 \Pn2 .
Then min-cut(G) ≤ min-cut(G) ≤ min-cut(Gn ). These inequalities can be shown
by the same argument as in the proof of Theorem 5. By construction, the max
flow on Gn is feasible on G, and therefore also optimal on G. Hence, by duality
min-cut(G) = min-cut(Gn ) which implies min-cut(G) = min-cut(Gn ).

15
There is a lot of redundancy in this algorithm. It is not necessary to compute
the max-flow from scratch in each iteration, especially in the augmenting paths
framework. Rather, starting with the max flow in Gi , flow can be pulled back
along s − t paths passing through vertices vp,1 or vp,2 for p ∈ P01 ∪ P02 until it
becomes feasible in graph Gi+1 . With such an initial flow, only a few augmenting
paths are required to find the max flow on Gi+1 . Since P 1 and P 2 are small
subsets of P, and P01 ∪P02 are small subsets of P 1 ∪P 2 , the cost of this algorithm
is negligible.

5 Unknown constant values, algorithm

The algorithm presented in the last sections minimizes Ed (c, φ1 , φ2 ) with respect
to φ1 , φ2 for a fixed c. Vice versa, for a fixed φ1 , φ2 the values c minimizing
Ed (c, φ1 , φ2 ) are given by the average intensity in each region

Ωi
u0 dx
ci = , i = 1, ..., n (35)
Ωi
dx

We want an algorithm to minimize both with respect to φ and c. As in [2], this

is achieved by combining the two above results in the following iterative descent
algorithm
Algorithm 2:
Estimate initial values c0 , set l = 0.
while( ||cl − cl−1 || > tol)
Use graph cuts to estimate φ from

φ = arg minφ̃1 ,φ̃2 Ed (cl , φ̃1 , φ̃2 ). (36)

Update cl+1 according to equation (35).

Update l ← l + 1
Note that no initialization of the level set functions are required. Only the
values c0 need to be initialized, which can be achieved very efficiently by the
isodata algorithm [32]. In our experiments, convergence was reached in around
5-15 iterations. It must be noted that this algorithm is no longer guaranteed
to find the global minima. Theoretically, it may get trapped in a local minima
close to the initial values c0 . However, in practice it is usually rather insensitive
to initialization.

6 Numerical results
Numerical examples are shown in Figure 4 - 7, where we have used the power
β = 2 in the data term. The constant values {ci }4i=1 are estimated by running
Algorithm 2 until convergence (6-10 iterations). During each iteration, the
energy minimization problem was submodular.

16
 (a)

Figure 4: L2 data fidelity. From left to right: input, level set method gradient
descent, our approach, alpha expansion/alpha beta swap.

(a)

Figure 5: Level set method: (a) bad initialization, (b) result.

In the relatively simple image in Figure, the level set method finds a good
local minima. If the initialization is bad, the level set method gets stuck in an
inferior local minima also for this simple image as shown in Figure 5. White
points indicate the zero level set of φ1 and dark points indicate the zero level
set of φ2 .
More difficult images are presented in Figure 6 - 7. The L2 data fidelity
term has been used (β = 2) and the different methods are compared by keeping
the same constant values {c∗i }4i=1 and regularization parameter ν fixed, while
minimizing in terms of the regions. One can clearly see the advantages of the
global approach over earlier approaches.

6.1 Experiments on L2 data fitting term: submodularity

In Section 3 we gave theoretical insights on how submodularity of the energy
function was related to the distribution of the values ci , i = 1, ..., 4. It was
shown that the condition becomes less strict as β increases. In this section we
demonstrate that for L2 data fitting term (β = 2 in (3)), the energy function
is submodular in practice. The L2 norm tolerates rather uneven distributions
of ci , i = 1, ..., 4. In addition, the parameters ci i = 1, ..., 4 minimizing the
energy function are not expected to get too clustered. To verify this, we have
run Algorithm 2 for optimizing the parameters ci , i = 1, ..., 4 on all images from
the data base [1]. For all the experiments, the submodularity condition was

17
 (a) (b)

(c) (d) (e)

Figure 6: Experiment 2: (a) Input, (b) gradient descent, (c) our approach, (d)
alpha expansion, (e) alpha-beta swap.

(a) (b) (c)

(d) (e) (f)

Figure 7: Experiment 3: (a) Input image, (b) ground truth, (c) gradient descent,
(d) our approach, (e) alpha expansion, (f) alpha-beta swap.

18
 (a) Experiment 5

Figure 8: L1 data fidelity. Note that the constant values of c1 , c2 , c3 are very
close to each other compared to c4 . From left to right: input image, set of pixels
P 1 ∪ P 2 where the submodular condition was not satisfied, set of pixels where
residual criterion (34) is not satisfied (empty set), output image.

satisfied during each iteration of the algorithm.

6.2 Experiments on non-submodular energy minimization

The purpose of this section is to demonstrate the methods for minimizing non-
submodular energy functions. For that reason, we will use the L1 data fidelity
term and fix the values ci , i = 1, ..., 4 in such a way that the submodular
condition (18) is violated. Figure 8 shows such an example, which is a modified
version of the example in Figure 4, where the average intensities values of 3 of
the objects are close compared to the 4th object. Some more natural examples
are shown in Figure 9 and 10. Subfigures (b) show the set of pixels p ∈ P1 ∪ P2 ,
where the submodular condition was violated. Subfigures (c) show the set of
pixels where the residual capacity conditions (34) were violated, which is the
empty set in all cases. Therefore, the solutions obtained by the cut on the
graphs G are also global solutions to the original problems.
If the regularization parameter ν is set extremely high, the residual capacity
condition (34) may also be violated at a small set of the pixels. Two such
examples are shown in Figure 11 and 12. As we see from subfigures (c), the
residual capacity condition is only violated at a small set of the pixels. By
applying the refinement algorithm from Section 4.2, the exact global solution
can be obtained in two iterations, as shown in subfigures (c)-(e).

7 Conclusions
We have developed a global minimization method for the multiphase Chan-Vese
model of image segmentation based on graph cuts. Numerical experiments also
demonstrated superior efficiency of the new approach over gradient descent.
It was shown that the energy function was submodular provided the average
intensity values of each region was sufficiently evenly distributed. The strictness
on of the condition depended on the data term. For Lp data term with p ≥ 2, the
condition was satisfied in all our experiments. For L1 data term the condition

19
 (a) (b)

(c) (d)

Figure 9: Segmentation with L1 data fidelity (β = 1): (a) Input; (b) Set of
pixels P 1 ∪ P 2 where the submodular condition was not satisfied; (c) Set of
pixels where residual capacity criterion (34) was not satisfied (empty set); (d)
Output.

(a) (b) (c) (d)

Figure 10: Segmentation with L1 data fidelity (β = 1): (a) Input; (b) Set of
pixels P 1 ∪ P 2 where the submodular condition was not satisfied; (c) Set of
pixels where residual capacity criterion (34) was not satisfied (empty set); (d)
Output.

20
 (a) (b) (c)

(d) (e) (f)

Figure 11: Segmentation with L1 data fidelity (β = 1) and very high regular-
ization ν: (a) Input image; (b) Set of pixels P 1 ∪ P 2 where the submodular
condition was not satisfied; (c) Set of pixels where the residual capacity condi-
tion (34) was violated; (c) -(e) Set of pixels where the weights on graph Gi differ
from the weights on the graph Gi−1 , i = 1, ..., 3; (f) Output (global solution).

was more easily violated. Algorithms for minimizing non-submodular functions

were developed, which most often computed global solution, but could not be
proved to always do so.
In this work, we have restricted our attention to four (or less) phases. The
results can be generalized to more phases by using more level set functions.
For m level set functions, m vertices in the graph will be associated to each
grid point. Since the data term then would involve interactions between m
binary variables, we expect submodularity to be more restrictive. We plan to
investigate how submodularity is related to the constant values in these cases,
and extend the non-submodular algorithm to this setting. On the other hand,
four phases suffices in theory to segment any 2D image by the four color theorem.
Therefore, algorithms can alternatively be designed to take advantage of this,
which makes extensions to more than four phases unnecessary.

A Proofs of Prop 1-4

A.1 Proof of Prop 1
c1 −c2 c4 −c3
Proof. Let 2 ≤I ≤ 2 . Then

|c2 − I|β ≤ |c1 − I|β and |c3 − I|β ≤ |c4 − I|β ,

21
 (a) (b) (c)

(d) (e) (f)

Figure 12: Segmentation with L1 data fidelity (β = 1) and very high regular-
ization ν: (a) Input image; (b) Set of pixels P 1 ∪ P 2 where the submodular
condition was not satisfied; (c) Set of pixels where the residual capacity condi-
tion (34) was violated; (c) -(e) Set of pixels where the weights on graph Gi differ
from the weights on the graph Gi−1 , i = 1, ..., 3; (f) Output (global solution).

22
for any β ≥ 1. Therefore, adding these two inequalities

|c2 − I|β + |c3 − I|β ≤ |c1 − I|β + |c4 − I|β .

A.2 Proof of Prop 3

c1 −c2 c4 −c3 c1 −c2
When 2 ≤I ≤ 2 , the result follows from Prop (2). Consider I < 2 ,
then
|I − c1 |β ≤ |I − c2 |β ≤ |I − c3 |β ≤ |I − c4 |β . (37)
Together with (10), this implies

0 < |I − c2 |β − |I − c1 |β ≤ |I − c4 |β − |I − c3 |β . (38)

Therefore, there exists θ1 ≥ θ2 > 1 such that

|I − c4 |β = θ1 |I − c3 |β and |I − c2 |β = θ2 |I − c1 |β . (39)

For α ≥ β

|I − c4 |α = θ1α−β |I − c3 |α and |I − c2 |α = θ2α−β |I − c1 |α , (40)

where θ1α−β ≥ θ2α−β > 1, hence

1 1
|I −c2 |α +|I −c3 |α = θ2α−β |I −c1 |α + |I −c4 |α ≤ θ1α−β |I −c1 |α + |I −c4 |α
θ1α−β θ1α−β
1
≤ θ1α−β |I − c1 |α + |I − c4 |α ,
θ1α−β
where the last inequality follows from |I − c 1 |α
≤ |I − c4 |α . Exactly the same
argument can be used to show Prop 3 when I > c3 −c
2 .
4

A.3 Proof of Prop 4

Proof. Assume first I > c3 , then

|c1 − I| > |c2 − I| > |c3 − I|

Therefore, there exists a θ > 1 s.t.

|I − c1 | = θ |c2 − I|.

Pick BI1 ∈ N s.t.

θβ ≥ 2, ∀β ≥ B.
Then

|c1 − I|β + |c4 − I|β ≥ |c1 − I|β ≥ 2|c2 − I|β > |c2 − I|β + |c3 − I|β . ∀β ≥ BI1

23
If I < c2 , then
|c4 − I| > |c3 − I| > |c2 − I|
and thus the same argument can be used to show there exists BI2 ∈ N such that

|c4 − I|β + |c1 − I|β > |c2 − I|β + |c3 − I|β , ∀β ≥ BI2 .

In case c2 ≤ I ≤ c3 , the existence of such a B was proved in Prop 1, e.g. B = 1.

Therefore the condition (10) is satisfied for any I ∈ [0, L] by choosing β ≥
B = maxI∈[0,L] max{BI1 , BI2 }.

A.4 Proof of Prop 2

We will show the condition holds for β = 1, which by Prop (3) implies it holds
for all β ≥ 1. Observe that if c1 , c2 and c3 , c4 are equidistant it follows that
c1 + c4 = c2 + c3 . For I < c2

|I − c2 | + |I − c3 | = (c2 − I) + (c3 − I) = −2I + (c2 + c3 )

= −2I + (c1 + c4 ) = (c1 − I) + (c4 − I) ≤ |I − c1 | + |I − c4 |.

For I ≥ c3

|I − c2 | + |I − c3 | = (I − c2 ) + (I − c3 ) = 2I − (c2 + c3 )

= 2I − (c1 + c4 ) = (I − c1 ) + (I − c4 ) ≤ |I − c1 | + |I − c4 |.

References
[1] http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/.
[2] E. Bae and X. Tai, Graph cut optimization for the piecewise constant
level set method applied to multiphase image segmentation, in Scale Space
and Variational Methods in Computer Vision, Springer, 2009, pp. 1–13.
[3] Y. Boykov and G. Funka-Lea, Graph cuts and efficient n-d image
segmentation, Int. J. Comput. Vision, 70 (2006), pp. 109–131.
[4] Y. Boykov and V. Kolmogorov, An experimental comparison of min-
cut/max-flow algorithms for energy minimization in vision, in Energy Min-
imization Methods in Computer Vision and Pattern Recognition, 2001,
pp. 359–374.
[5] , Computing geodesics and minimal surfaces via graph cuts, in ICCV
’03: Proceedings of the Ninth IEEE International Conference on Computer
Vision, Washington, DC, USA, 2003, IEEE Computer Society, p. 26.
[6] Y. Boykov, V. Kolmogorov, D. Cremers, and A. Delong, An
integral solution to surface evolution pdes via geo-cuts, ECCV06, (2006),
pp. 409–422.

24
 [7] Y. Boykov, O. Veksler, and R. Zabih, Fast approximate energy min-
imization via graph cuts, in ICCV (1), 1999, pp. 377–384.
[8] X. Bresson and T. Chan, Non-local unsupervised variational image
segmentation models, UCLA, Applied Mathematics, CAM-report-08-67,
(2008).
[9] A. Chambolle, Total variation minimization and a class of binary mrf
models, in Energy Minimization Methods in Computer Vision and Pattern
Recognition, Springer Berlin / Heidelberg, 2005, pp. 136–152.
[10] T. Chan and L. Vese, Active contours without edges, IEEE Image Proc.,
10, pp. 266-277, 2001.
[11] J. Darbon, A note on the discrete binary mumford-shah model, Proceed-
ings of Computer Vision/Computer Graphics Collaboration Techniques,
(MIRAGE 2007), LNCS Series vol. 44182, (March 2007), pp. 283–294.
[12] J. Darbon and M. Sigelle, Image restoration with discrete constrained
total variation part i: Fast and exact optimization, J. Math. Imaging Vis.,
26 (2006), pp. 261–276.
[13] , Image restoration with discrete constrained total variation part ii:
Levelable functions, convex priors and non-convex cases, J. Math. Imaging
Vis., 26 (2006), pp. 277–291.
[14] A. Dervieux and F. Thomasset, A finite element method for the simula-
tion of a Rayleigh-Taylor instability, in Approximation methods for Navier-
Stokes problems (Proc. Sympos., Univ. Paderborn, Paderborn, 1979),
vol. 771 of Lecture Notes in Math., Springer, Berlin, 1980, pp. 145–158.
[15] S. Geman and D. Geman, Stochastic relaxation, gibbs distributions, and
the bayesian restoration of images, in Readings in uncertain reasoning, Mor-
gan Kaufmann Publishers Inc., San Francisco, CA, USA, 1990, pp. 452–472.
[16] D. M. Greig, B. T. Porteous, and A. H. Seheult, Exact maximum
a posteriori estimation for binary images, Journal of the Royal Statistical
Society, Series B, (1989), pp. 271–279.
[17] H. Ishikawa and D. Geiger, Segmentation by grouping junctions, in
CVPR ’98: Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, Washington, DC, USA, 1998,
IEEE Computer Society, pp. 125–131.
[18] V. Kolmogorov and C. Rother, Minimizing nonsubmodular functions
with graph cuts-a review, IEEE Trans. Pattern Anal. Mach. Intell., 29
(2007), pp. 1274–1279.
[19] V. Kolmogorov and R. Zabih, Multi-camera scene reconstruction via
graph cuts, in ECCV ’02: Proceedings of the 7th European Conference on
Computer Vision-Part III, London, UK, 2002, Springer-Verlag, pp. 82–96.

25
[20] , What energy functions can be minimized via graph cuts?, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 26 (2004),
pp. 147–159.
[21] N. Komodakis, G. Tziritas, and N. Paragios, Fast, approximately
optimal solutions for single and dynamic mrfs, Computer Vision and Pat-
tern Recognition, 2007. CVPR ’07. IEEE Conference on, (17-22 June 2007),
pp. 1–8.
[22] D. F. L. Ford, Flows in networks, Princeton University Press, 1962.
[23] J. Lellmann, J. Kappes, J. Yuan, F. Becker, and C. Schnorr,
Convex multi-class image labeling by simplex-constrained total variation, in
SSVM09, 2009, pp. 150–162.
[24] J. Lie, M. Lysaker, and X. Tai, A binary level set model and some
applications to mumford-shah image segmentation, IEEE Transactions on
Image Processing, 15 (2006), pp. 1171–1181.
[25] , A variant of the level set method and applications to image segmen-
tation, Math. Comp., 75 (2006), pp. 1155–1174 (electronic).
[26] J. Lie, M. Lysaker, and X.-C. Tai, A variant of the level set method and
applications to image segmentation, UCLA, Applied Mathematics, CAM-
report-03-50, (2003).
[27] D. Mumford and J. Shah, Optimal approximation by piecewise smooth
functions and associated variational problems, Comm. Pure Appl. Math,
42, 42 (1989), pp. 577–685.
[28] M. Nikolova, S. Esedoglu, and T. F. Chan, Algorithms for find-
ing global minimizers of image segmentation and denoising models, SIAM
Journal on Applied Mathematics, 66 (2006), pp. 1632–1648.
[29] S. Osher and J. Sethian, Fronts propagating with curvature dependent
speed: algorithms based on hamilton-jacobi formulations, J. Comput. Phys.,
79 (1988), pp. 12–49.
[30] T. Pock, A. Chambolle, H. Bischof, and D. Cremers, A convex
relaxation approach for computing minimal partitions, in IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), Miami, Florida,
2009. to appear.
[31] C. Rother, S. Kumar, V. Kolmogorov, and A. Blake, Digital
tapestry, in CVPR ’05: Proceedings of the 2005 IEEE Computer Soci-
ety Conference on Computer Vision and Pattern Recognition (CVPR’05) -
Volume 1, Washington, DC, USA, 2005, pp. 589–596.
[32] F. R. D. Velasco, Thresholding using the ISODATA clustering algorithm,
IEEE Trans. Systems Man Cybernet., 10 (1980), pp. 771–774.

26
[33] L. A. Vese and T. F. Chan, A new multiphase level set framework for
image segmentation via the mumford and shah model, International Journal
of Computer Vision, 50 (2002), pp. 271–293.
[34] C. Zach, D. Gallup, J.-M. Frahm, and M. Niethammer, Fast global
labeling for real-time stereo using multiple plane sweeps, in Vision, Modeling
and Visualization Workshop (VMV), 2008.
[35] N. E. Zehiry, S. Xu, P. Sahoo, and A. Elmaghraby, Graph cut
optimization for the mumford-shah model, in Proceedings of the Seventh
IASTED International Conference visualization, imaging and image pro-
cessing, Springer Verlag, 2007, pp. 182–187.

27