.Petr Hliněný, FI MU Brno, 2019 1 / 29 FI: MA010: Cuts and Network Flows
4 Graph Cuts and Network Flows
Yet another area of rich applications of graphs (digraphs) deals with so called flow networks
and “commodity flows” in them.
✷
Brief outline of this lecture
• More on graph connectivity – vertex and edge cuts in graphs.
• Flow networks, admissible flows and network cuts, circulations.
• Finding a maximum flow via augmenting of residual paths.
• Extensions, and applications in connectivity, SDR and vision.
.Petr Hliněný, FI MU Brno, 2019 2 / 29 FI: MA010: Cuts and Network Flows
4.1 Connectivity and Cuts
Recall from Lecture 2:
• A graph G is edge-k-connected, k > 1, if G stays connected even after removal
of any subset of ≤ k − 1 edges. ✷
• A graph G is vertex-k-connected, k > 1, if G has more than k vertices and G
stays connected even after removal of any subset of ≤ k − 1 vertices. ✷
• Menger’s theorem: A graph G is k-connected iff there exist ≥ k internally
disjoint paths between any pair of vertices (the paths may share only their ends).
s s
ss
s s❢ ❢
❢
❢
Remark: The obstacles to high connectivity in graphs are small so-called cuts.
vertex/ edge cut = hranový / vrcholový ˇrez
minimal cut = minimální ˇrez
.Petr Hliněný, FI MU Brno, 2019 3 / 29 FI: MA010: Cuts and Network Flows
Edge / vertex cuts
An s–t path in a graph G is a path in G with the ends s, t ∈ V (G).
s ss t
Definition: Let G be a graph and s, t ∈ V (G). An edge set F ⊆ E(G) is an s–t edge
cut in G if the subgraph G \ F (deletion of the edges F from G) has no s–t path.✷
In other words, if s and t belong to different conn. components of the subgraph G \ F.
✷
s ss t
s
s
❢
❢
Similarly, a vertex set X ⊆ V (G) is an s–t vertex cut in G if the subgraph G \ X
(deletion of the vertices X from G) has no s–t path. ✷
An s–t cut is called minimal if no proper subset of it is an s–t cut again. ✷
If s–t are implic. known, or irrelevant, then we shortly say an edge / vertex cut in G.
.Petr Hliněný, FI MU Brno, 2019 4 / 29 FI: MA010: Cuts and Network Flows
Cuts or bipartitions?
Again (unfortunately), there is a bit of confusion in terminology. . .
• Our edge cut is any subset of edges whose removal disconnects some two
vertices.✷
• Sometimes, the following is used as a definition of an edge cut:
A set F ⊆ E(G) is called (here) a b-cut in G if it is of the form (“bipartition”)
F = uv : uv ∈ E(G) ∧ u ∈ V1 ∧ v ∈ V2 where V1 ∩ V2 = ∅, V1 ∪ V2 = V (G).
V1 V2
✷
• Not every edge cut is a b-cut, and not every b-cut is an edge cut. ✷However, it is true:
– Every minimal nonempty edge cut is a b-cut.
– Every minimal nonempty b-cut is an edge cut.
bridge = most, cutvertex = artikulace, disconnected graph = nesouvislý graf
.Petr Hliněný, FI MU Brno, 2019 5 / 29 FI: MA010: Cuts and Network Flows
Small cuts and blocks
Definition: A bridge in a graph is a minimal edge cut consisting of one edge.
A cutvertex in a graph is a minimal vertex cut consisting of one vertex.
s s
ss
s ss s
s s
ss
s ss❢
✷
Fact: A connected graph G is not 2-connected iff G has a cutvertex or G ≃ K1, K2.✷
A graph G on |V (G)| > k vertices is not k-connected iff G has a vertex cut of size
k − 1 or less. (Note that a disconnected graph has a cut of size 0.) ✷
Definition: A block in a graph G is a maximal (by inclusion) 2-connected subgraph
of G. Moreover, a vertex with no edges and an edge not contained in any larger block
of G are also called blocks of G. ✷
Proposition 4.1. A subgraph H ⊆ G is a block if, and only if, H is max. connected
(among subgraphs of G) with the property that H contains no cutvertex (of H itself).
... by contradiction = d˚ukaz sporem
.Petr Hliněný, FI MU Brno, 2019 6 / 29 FI: MA010: Cuts and Network Flows
An excersise: the structure of graph blocks
s
s
s s
s
s
s
Proposition 4.2. Let G be a graph and B1, B2, . . . , Bk be the blocks of G. If Bi ∩
Bj = ∅, then Bi ∩ Bj = {c} where c is a cutvertex of G. ✷
Proof by means of contradiction: Let Bi∩Bj ⊇ {c, d} where c = d. Since Bj = {c, d},
we have that Bj is 2-connected by the definition, and so there exists a path P ⊆ Bj \cd
with the ends c, d. ✷Let B+
:= Bi ∪ P ⊆ G. Since Bi is 2-connected, so is B+
(Thm. 2.10, “adding an ear P”). However, this contradicts the definition of a block;
Bi already was a maximal 2-connected subgraph of G. ✷
The same argument proves the last part, that c ∈ Bi ∩ Bj is a cutvertex, too. Assume
not, then for some neighbours u, v of c such that u ∈ V (Bi) and v ∈ V (Bj), there
exists a u–v path Q ⊆ G\c and B+
:= Bi ∪(Q+vc) contradicts maximality of Bi.✷✷
Corollary 4.3. The bipartite “incidence graph” between the blocks and the cutvertices
of a graph G is a forest (it contains no cycles).
.Petr Hliněný, FI MU Brno, 2019 7 / 29 FI: MA010: Cuts and Network Flows
Menger’s theorem reformulated
Recall. . .
• Menger’s theorem: A graph G is k-connected iff there exist ≥ k internally
disjoint paths between any pair of vertices (the paths may share only their ends).
Theorem 4.4. Let G be a graph and s, t ∈ V (G). There exist k internally disjoint
s–t paths in G if, and only if, there is no s–t vertex cut in G of size less than k. ✷
s ss t
s
s
❢
❢
A sketch: the “only if” direction is obvious, and for the “if” direction, ✷
we will generalize the whole setting to “weighted connectivity” next. . .
flow network = sít’, source = zdroj, sink = stok, capacity = kapacita
.Petr Hliněný, FI MU Brno, 2019 8 / 29 FI: MA010: Cuts and Network Flows
4.2 Digraphs as Flow Networks
Flow networks present a convenient “weighted generalization” of the concept of graph
connectivity and cuts, with long history of research and many practical applications. ✷
Definition 4.5. A flow network is a quadruple ¯G = (G, z, s, w) such that
– G is a digraph (it is important to have oriented edges),
– the vertices z ∈ V (G), s ∈ V (G) are the source and the sink, respectively,
– and w : E(G) → R+
∪ {∞} is a positive weighting of the arcs (edges) of G,
these weights are called edge capacities.
s s
ss
3
1
4
5
1
5
3
2
2
4
2
z s
✷
Remark: In a real world, more than one source or sink may exist in a flow network, but that
is not a problem — we simply create a single artificial “supersource” and draw arcs from it to
all the real sources (even with source capacities), and the same with a “supersink”.
network flow = tok v síti
admissible = pˇrípustný, capacity constraints = kapacitní omezení, flow conservation = zachování toku
.Petr Hliněný, FI MU Brno, 2019 9 / 29 FI: MA010: Cuts and Network Flows
Notation: For simplicity, we shall write e → v to mean that an arc e “points to” (has
its head in) the vertex v, and e ← v analogously for e “leaving” (having tail in) v. ✷
Definition 4.6. A network flow, in a flow network ¯G = (G, z, s, w),
is an assignment f : E(G) → R+
0 satisfying (we say f is admissible)
– ∀e ∈ E(G) : 0 ≤ f(e) ≤ w(e), and (capacity constraints)
– ∀v ∈ V (G), v = z, s :
e→v
f(e) =
e←v
f(e). (flow conservation) ✷
The value (size) of a flow f is the quantity f =
e←z
f(e) −
e→z
f(e). ✷
Notation: The flow value F and the capacity C of an arc in a picture of a network
will be shortly denoted by F/C, respectively.
s s
ss
3/3
0/1
2/4
3/5
0/1
1/2
1/2
3/4
2/2
2/5
0/3
z s
circulation = cirkulace (v síti)
.Petr Hliněný, FI MU Brno, 2019 10 / 29 FI: MA010: Cuts and Network Flows
Circulations
Definition: A flow in a flow network ¯G = (G, z, s, w), satisfying the flow conservation
constraint at all the vertices including the source and the sink z, s, is called a circulation.
The source and the sink hence become irrelevant for circulations. ✷
Fact: There is a one-to-one correspondence between a flow f in a flow network ¯G and
the following circulation in the enhanced network ¯G + (s, z):
f
z s
• simply add the “reverse” arc (s, z) assigned flow value f and capacity +∞.
boundary surplus = pˇrebytek na hranici
.Petr Hliněný, FI MU Brno, 2019 11 / 29 FI: MA010: Cuts and Network Flows
Boundary property of flows
Notation: For ¯G and U ⊆ V (G), let e → U mean that an arc e points from V (G)\U
towards U, and e ← U mean that an arc e points from U towards V (G) \ U. ✷
Definition: Let ¯G be a flow network and U ⊆ V (G). The boundary surplus of a
flow f on U is
e←U
f(e) −
e→U
f(e) .✷
In this picture, the surplus on U is 5:
s s
ss
3/3
0/1
2/4
4/5
0/1
0/2
1/2
4/4
1/2
2/5
0/3
U
z s
✷
Fact: The boundary surplus of f on {z} is exactly the flow size f by the definition.
We aim to extend this finding to other suitable U’s.
.Petr Hliněný, FI MU Brno, 2019 12 / 29 FI: MA010: Cuts and Network Flows
Lemma 4.7. Let ¯G be a flow network and ∅ = U ⊆ V (G). If f is a circulation in ¯G,
then the boundary surplus of f on U is always 0.
Proof: We use induction on |U|. For U = {x}, the claim is just the flow conservation
constraint at x. ✷
Consider now arbitrary U ⊆ V (G) where |U| ≥ 2 and y ∈ U. By the induction
assumption, the claim holds for U′
= U \ {y}. Now we compute
e←U
f(e) =


e←U′
f(e) −
e←U′ ∧ e→y
f(e)

 +


e←y
f(e) −
e←y ∧ e→U′
f(e)

 ✷
and similarly
e→U
f(e) =


e→U′
f(e) −
e→U′ ∧ e←y
f(e)

 +


e→y
f(e) −
e→y ∧ e←U′
f(e)

 ✷.
Therefore, the new surplus is
e←U
f(e) −
e→U
f(e) =
e←U′
f(e) −
e→U′
f(e) +
e←y
f(e) −
e→y
f(e) = 0 + 0.
✷
.Petr Hliněný, FI MU Brno, 2019 13 / 29 FI: MA010: Cuts and Network Flows
Where to measure flow size
Lemma 4.7. (repeated) Let ¯G be a flow network and ∅ = U ⊆ V (G). If f is
a circulation in ¯G, then the boundary surplus of f on U is always 0.
Corollary 4.8. Let ¯G = (G, z, s, w) be a flow network and f an admissible flow in ¯G.
For any U ⊆ V (G) such that z ∈ U ∋ s (U “separating” z from s), the boundary
surplus of f on U is always the same, equal to f . ✷
Proof: Turn f into a circulation f′
in G+sz by letting f′
≡ f on G and f′
(sz) = f ,
and apply Lemma 4.7. U
f
z s
Then, for any such U, it is sz → U and so f′
(sz) is accounted for in the boundary
surplus of f′
on U, which is 0. The surplus of f on U hence equals f′
(sz) = f . ✷
Vˇeta o maximálním toku a minimálním ˇrezu.
.Petr Hliněný, FI MU Brno, 2019 14 / 29 FI: MA010: Cuts and Network Flows
4.3 The Max-flow Min-cut Theorem
Edge cuts in graphs have a natural weighted generalization into flow networks. . .
Definition 4.9. A cut in a flow network ¯G = (G, z, s, w)
is a subset of edges (arcs) C ⊂ E(G) such that there is no z → s directed path in G
completely avoiding C (i.e., s is not reachable from z in G \ C). ✷
The capacity (size) of a cut C is the sum of the capacities of arcs in C, i.e.,
C =
e∈C
w(e).✷
The most important result in this area is the following:
Theorem 4.10. (Ford–Fulkerson) In any flow network ¯G, there exists an admissible
flow of size r ∈ R+
if, and only if, there is no cut in ¯G of capacity less than r. ✷
Maximum possible flow size
=
minimum cut capacity.
.Petr Hliněný, FI MU Brno, 2019 15 / 29 FI: MA010: Cuts and Network Flows
An example
What is a cut of the least capacity in this example?
Is it unique?
s s
ss
3
4
4
1
1
4
2
1
1
C = 5C′
= 5
z s
We can see two cuts C and C′
of the same size 5, and no smaller one. . .
good characterizations = dobrá charakterizace, obstacle = pˇrekážka
.Petr Hliněný, FI MU Brno, 2019 16 / 29 FI: MA010: Cuts and Network Flows
On the flow–cut duality
Theorem 4.10 (repeated). In any flow network ¯G, there exists an admissible flow
of size r ∈ R+
if, and only if, there is no cut in ¯G of capacity less than r.
How to read this theorem? ✷
• If one looks for a certificate that a flow is maximum possible, then it always
suffices to exhibit a cut of the same value. ✷
• Likewise, to certify minimality of a cut, one exhibits a flow of the same value. ✷
• Nice properties of this kind are commonly called good characterizations (of a
problem)—one can certify optimality of a solution by giving an obvious obstacle.✷
Such as, in our case, the following:
Proposition 4.11. In any ¯G, for any adm. flow f and any cut C, it holds f ≤ C .✷
Proof: Let U be the set of vertices of G reachable from z in G \ C, where s ∈ U. Let
FU = {e : e ← U}. Then FU ⊆ C by the definition of U. By Corollary 4.8,
f =
e∈FU
f(e) −
e→U
(e) ≤
e∈FU
w(e) ≤ C .
✷
.Petr Hliněný, FI MU Brno, 2019 17 / 29 FI: MA010: Cuts and Network Flows
4.4 Finding the Maximum Flow
A question: is the following flow maximum possible?
s s
ss
3/3
0/1
2/4
3/5
0/1
1/2
1/2
3/4
2/2
2/5
0/3
✷
3=2/4
4=3/5
4=3/4
1=0/3
z s
Now it is, of size 6, and we have got a cut of capacity 6 as well. ✷
Fact: There exist quite simple and fast algorithms to find the maximum flow (and the
minimum cut at the same time) in a given flow network.
– These simple algorithms are based on an idea to saturate “residual z–s paths”.
.Petr Hliněný, FI MU Brno, 2019 18 / 29 FI: MA010: Cuts and Network Flows
Problem formulation
Problem 4.12. The Max–Flow problem
Given a flow network ¯G = (G, z, s, w), the task is to find a flow f in ¯G from z to s
such that the flow size f is maximized (among all admissible flows in ¯G). ✷
Where can one find the max-flow problem?
• in transportation (or distribution) networks of goods, electricity, etc.
• in pipe networks (water, gas, oil, sewerage, etc.) ✷
• in IP packet routing (real-time transmission of large data), ✷
• in various “matching” or “representatives” problems, ✷
• in computer vision – as image segmentation (the min-cut).
residual path = nenasycená (zbytková) cesta, residual capacity = zbytková kapacita
.Petr Hliněný, FI MU Brno, 2019 19 / 29 FI: MA010: Cuts and Network Flows
Residual and Augmenting Paths
Definition: Consider a flow network ¯G and an admissible flow f in it.
A residual z– x path (in ¯G w.r.t. f and the source z) is an undirected path in G from
the source z to any vertex x, i.e. a sequence of adjacent edges e1, e2, . . . , em; ✷
– such that f(ei) < w(ei) if ei is directed away from z,
– and f(ei) > 0 if ei is directed towards z (“backwards”).
s s s s
3/4 1/2 1/1 2/3 2/4
+1 +1 +1 +2 +2residual capacities: > 0
z s
The quantity w(ei) − f(ei), or f(ei), resp., is called the residual capacity of edge ei.✷
The background idea is as follows.
• A residual z–s path has strictly positive residual capacity ε > 0, and so one can
“push” an additional ε amount of flow from the source z to the sink s. ✷
• However, what does “pushing flow against an arc” mean??? ✷
Actually, we stop a (bit of) “returning” flow, and send new amount of flow instead
of it. (Imagine stoppig a crowd of returning people and thus making more room.)
augmentation = zvˇetšení
.Petr Hliněný, FI MU Brno, 2019 20 / 29 FI: MA010: Cuts and Network Flows
Method 4.13. Maximizing flow via residual paths.
The overall idea is really simple; ✷
one should repeatedly augment (meaning to enlarge) the current flow by adding to it
along existing residual paths. . .
How to augment a residual path:
s s s s
3/4 1/2 1/1 2/3 2/4
+1 +1 +1 +2 +2residual capacities: > 0
z s
✷
min. residual capacity r = 1 > 0 ❀ augmenting the current flow by +1:
s s s s
4/4 2/2 0/1 1/3 3/4
z s
✷
Fact: Augmenting (enlarging) an admissible flow by the minimal residual capacity of
a residual z– s path results in an admissible flow, again.
.Petr Hliněný, FI MU Brno, 2019 21 / 29 FI: MA010: Cuts and Network Flows
A simple Residual Path Algorithm
s s
ss
3
4
4
1
1
4
2
1
1
z s
Algorithm 4.14. Ford–Fulkerson’s for network flows.
input ← a flow network ¯G = (G, z, s, w);
flow f ≡ 0;
repeat {
Search (BFS) the graph G to find the set U of those vertices
reachable from the source z along residual paths;
if ( s ∈ U ) {
P = any residual z–s path in ¯G (this P then called an augmenting path);
Augment (“enlarge”) f by the minimal residual capacity along P;
}
until ( s ∈ U );
output → a maximum flow f in ¯G;
output → a minimum cut in ¯G from U to V (G) − U.
.Petr Hliněný, FI MU Brno, 2019 22 / 29 FI: MA010: Cuts and Network Flows
Proof of Algorithm 4.14:
For any flow f and any cut C in ¯G, it holds f ≤ C . If the algorithm stops with a
flow f in ¯G and a cut C such that C = f , then it is clear that f is a maximum
flow in ¯G. (We have, however, not proved yet that the algorithm stops!) ✷
So to prove that whenever the algorithm stops with f, C, then f = C , we use the
following schematic picture (in which s does not belong to the reachable set U):
f(e) = w(e)
f(e) = 0
U
z s
✷
Since no further vertex than U is reachable along residual paths, every arc e leaving
U has full flow f(e) = w(e), and every arc e entering U has zero flow f(e) = 0.
Therefore;
e←U
f(e) −
e→U
f(e) =
e←U
f(e) =
e∈C
w(e) = C .
Finally, by Corollary 4.8, we have f = e←U f(e) − e→U f(e) = C , finishing
the proof. ✷
.Petr Hliněný, FI MU Brno, 2019 23 / 29 FI: MA010: Cuts and Network Flows
Basic Consequences
Algorithm 4.14 and its proof provides several interesting mathematical findings:
• Together with Prop. 4.11, it nearly(!) proves Theorem 4.10 (flow-cut duality); ✷
– what is missing, is a proof that the algorithm terminates, ✷
– and the latter is not obvious at all for arbitrary real capacities—there do
exist non-terminating (and even non-convergent) real examples. ✷
• If all the capacities are integers, then every step of Algorithm 4.14 deals with
integral residua and flows; ✷
– consequently, the algorithm must terminate under any circumstances,
– and the resulting flow will be integral as well. ✷
• For instance, if the edge capacities are all set to 1, one obtains the following.
For a graph and vertices s, t, there exist k edge-disjoint s–t paths iff there
is no s–t edge cut of size less than k. ✷
Compare this to Menger’s theorem, and to Section 4.5. . .
.Petr Hliněný, FI MU Brno, 2019 24 / 29 FI: MA010: Cuts and Network Flows
More precise formulations
. . . of the previous findings. . .
Theorem 4.15. (Edmonds–Karp)
If Algorithm 4.14 always chooses an augmenting path among residual paths of the
least length (measured by the number of edges), e.g., by the BFS, then the algorithm
is guaranteed to terminate after O(|V (G)| · |E(G)|) iterations.
Consequently, Theorem 4.10 is proved by this one. ✷
Proposition 4.16. If the edge capacities in a flow network ¯G are integral, then there
exists a maximum flow which is integral, too.
Algorithm 4.14 outputs such an integral flow in a finite number of steps.
.Petr Hliněný, FI MU Brno, 2019 25 / 29 FI: MA010: Cuts and Network Flows
4.5 Further Improvements and Extensions
More efficient flow algorithms
• Edmonds–Karp (Theorem 4.15):
BFS is used to search for an augmenting path, runtime O(|V (G)| · |E(G)|2
). ✷
• Dinitz:
– BFS is used to find all the shortest residual paths in ¯G, creating a “layered”
residual network.
– The layered network is then completely saturated in one run.
Only O(|V (G)|) iterations of the main cycle, runtime O(|V (G)|2
· |E(G)|). ✷
• MPM “Three Indians”:
Similar to [Dinitz], but a layered network is saturated faster, runtime O(|V (G)|3
).✷
• An advance note on a special “planar” case:
A minimum cut can be found as a shortest path in the dual graph (see Lecture 7).
.Petr Hliněný, FI MU Brno, 2019 26 / 29 FI: MA010: Cuts and Network Flows
Networks with lower capacities
In a flow network with lower capacities, in addition to the weight function w, there is
another weight function ℓ : E(G) → R+
0 giving the lower edge capacities.
A flow f is then admissible if ℓ(e) ≤ f(e) ≤ w(e) for every edge e of the network. ✷
Notice that an admissible flow may not exist in such a lower-capacitated network. ✷
Algorithm 4.17. Max-flow in a lower-capacitated network
The solution in a network ¯G = (G, z, s, w, ℓ) is found in two stages:
• First, an admissible circulation r is found in ¯G + sz, respecting both the lower
an upper bounds ℓ, w. This is done by finding a maximum flow in an artificial
network modelling the “surplus” of lower capacities at every vertex. . . ✷
• Second, a maximum flow h is found in a modified network ¯G′
= (G, z, s, w − r)
(Alg. 4.14); this network has only upper capacities w(e) − r(e) for an edge e. ✷
• The resulting flow is the sum f ≡ r + h.
.Petr Hliněný, FI MU Brno, 2019 27 / 29 FI: MA010: Cuts and Network Flows
Networks with vertex capacities
In a flow network with vertex capacities (retaining edge capacities as well), the capacity
function is w : E(G) ∪ V (G) → R+
.
The meaning of capacity constraints at the vertices for admissible flows is that the
total sum of incoming flow to any vertex x is not more than w(x).
(Differently applicable to the source or the sink, though. . . )✷
Algorithm 4.18. Max-flow in a vertex-capacitated network
Translate a vertex-capacitated network to an ordinary flow network via “doubling” the
capacitated vertices (replacing them with new arcs between the two copies), as follows:
s
s
5
43
5
2
a
b
z
s
❀
s
ss
s
5
4
3
5
2
a
b
z
s
Then solve the ordinary network with Algorithm 4.14.
.Petr Hliněný, FI MU Brno, 2019 28 / 29 FI: MA010: Cuts and Network Flows
Back to Menger’s theorem
Setting capacities of all vertices (except the terminals s, t) to 1, immediately proves:
Theorem 4.4. Let G be a graph and s, t ∈ V (G). There exist k internally disjoint
s–t paths in G if, and only if, there is no s–t vertex cut in G of size less than k. ✷
Systems of distinct representatives (SDR)
Definition: Let M1, M2, . . . , Mk be a collection of nonempty sets. A system of distinct
representatives (SDR) of the set family {M1, M2, . . . , Mk} is a sequence of pairwise
distinct elements (x1, x2, . . . , xk) such that xi ∈ Mi for i = 1, 2, . . ., k. ✷
Theorem 4.19. (Hall) Let {M1, M2, . . . , Mk} be a family of nonempty sets. Then
there exists a system of its distinct representatives if, and only if,
∀J ⊆ {1, 2, . . ., k} :
j∈J
Mj ≥ |J| ,
i.e., the union of any subfamily of these sets has at least that many elements as the
number of sets in it.
Necessity of Hall’s condition in this theorem is obvious, and its sufficiency can be proved by
an application of network flows again. . .
.Petr Hliněný, FI MU Brno, 2019 29 / 29 FI: MA010: Cuts and Network Flows
Flows in image segmentation
Yet another profitable application of flow networks lies in the computer vision area:
The basic image segmentation problem asks for decomposing a given image into a
foreground and a background.
• Let the input consists of a pixel matrix, each pixel carrying two values of likelihood
to be in the foreground and in the background. Additionally, separation penalties
are assigned to the neighbouring pairs of pixels. ✷
• The network is constructed by introducing a source z and a sink s such that the
arcs from z to the pixels have their capacities equal to the foreground likelihood,
and the arcs from the pixels to s have their capacities equal to the background
likelihood. Additional bidirectional arcs join the neighbouring pixel pairs.
• A minimal cut in this network then defines the separation between foreground
and background parts.