See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/246346467

"GrowCut" - Interactive Multi-Label ND Image Segmentation By Cellular

Automata

Article · November 2004

CITATIONS READS

430 1,280

2 authors, including:

Vadim Konushin
Lomonosov Moscow State University
26 PUBLICATIONS   600 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Third opinion View project

All content following this page was uploaded by Vadim Konushin on 07 July 2014.

The user has requested enhancement of the downloaded file.

 “GrowCut” - Interactive Multi-Label N-D Image Segmentation By Cellular
Automata
Vladimir Vezhnevets ∗ Vadim Konouchine
Graphics and Media Laboratory †
Faculty of Computational Mathematics and Cybernetics
Moscow State University,
Moscow, Russia.

Abstract 4. Performs multi-label image segmentation (the computation

time does not depend on the number of labels);
In this paper we describe a novel algorithm for interactive multi-
label segmentation of N-dimensional images. Given a small num- 5. Is extensible, allows constructing new families of segmenta-
ber of user-labelled pixels, the rest of the image is segmented au- tion algorithms with specific properties;
tomatically by a Cellular Automaton. The process is iterative, as
the automaton labels the image, user can observe the segmentation 6. Is truly interactive - the user observes the process of comput-
evolution and guide the algorithm with human input where the seg- ing the segmentation and is able to make modifications and
mentation is difficult to compute. In the areas, where the segmen- corrections at any time;
tation is reliably computed automatically no additional user effort
is required. Results of segmenting generic photos and medical im- Second, and most important, we would like to attract attention to
ages are presented. Our experiments show that modest user effort another possible view on segmentation problem, apart from graph
is required for segmentation of moderately hard images. theory. Cellular automata family is very rich and diverse, and our
hope is that by applying experience and knowledge accumulated in
Keywords: interactive image segmentation, graph cut, cellular this field to interactive image segmentation task, research commu-
automata, image editing, foreground extraction nity will be able to come up with some novel and ingenious solu-
tions.
1 Introduction
1.1 Related work
Image segmentation is an integral part of image processing appli-
cations like medical images analysis and photo editing. A wide In this section we briefly outline main features of current state-of-
range of computational vision algorithms can also benefit from ex- the-art interactive segmentation techniques. We informally divide
istence of reliable and efficient image segmentation technique. For these methods into two families - proposed for generic image edit-
instance, intermediate-level vision problems such as shape from sil- ing, and developed especially for medical images. In reality, most
houette, shape from stereo and object tracking could make use of of the methods can be successfully applied in both domains.
reliable segmentation of the object of interest from the rest of the
scene. Higher-level problems such as recognition and image index- Magic Wand - is a common selection tool for almost any image
ing can also make use of segmentation results in matching. editor nowadays. It gathers color statistics from the user-
Fully automated segmentation techniques are being constantly specified image point (or region) and segments (connected)
improved, however, no automated image analysis technique can image region with pixels, which color properties fall within
be applied fully autonomously with guaranteed results in general some given tolerance of the gathered statistics.
case. That is why semi-automatic segmentation techniques that al-
low solving moderate and hard segmentation tasks by modest effort Intelligent paint - is a region-based interactive segmentation tech-
on the part of the user are becoming more and more popular. nique, based on hierarchical image segmentation by tobog-
Several powerful techniques for interactive image segmentation ganing [Reese 1999]. It uses a connect-and-collect strategy to
have been proposed recently based on graph cuts [Boykov and Jolly define an objects region. This strategy uses a hierarchical to-
2001], [Rother et al. 2004] and random walker [Grady and Funka- bogganing algorithm to automatically connect image regions
Lea 2004]. They seem to significantly outperform earlier methods that naturally flow together, and a user-guided, cumulative
both by resulting segmentation quality and required user effort. cost-ordered expansion interface to interactively collect those
This paper’s contribution is twofold. First, we propose a new regions which constitute the object of interest. This strategy
interactive segmentation scheme, based on cellular automata, that coordinates human-computer interaction to extract regions of
has several favorable properties (see section 2 for details): interest from complex backgrounds using paint strokes with a
mouse.
1. Capable of solving moderately hard segmentation tasks (see
examples in this paper); Intelligent scissors - is a boundary-based method, that computes
2. Is easy in implementing and allows efficient parallel imple- minimum-cost path between user-specified boundary points
mentation; [Mortensen and Barrett 1998]. It treats each pixel as a graph
node and uses shortest-path graph algorithms for boundary
3. Works with images of any dimension N ≥ 1; calculation. A faster variant of region-based intelligent scis-
sors uses tobogganing for image oversegmentation and then
∗ e-mail: [email protected] treats homogenous regions as a graph nodes [Mortensen and
† www: http://graphics.cmc.msu.ru Barrett 1999].
Graph Cut is a combinatorial optimization technique, which was achieving better segmentation results with less user effort required,
applied by [Boykov and Jolly 2001] to the task of image seg- compared with other methods. One of the few drawbacks of the
mentation. The image is treated as a graph - each pixel is a graph-based methods is that they are not easily extended to multi-
graph node. For the case of two labels (i.e. object and back- label task and the other is that they are not very flexible - the only
ground) the globally optimal pixel labelling (with respect to tunable parameters are the graph weighting and cost function coef-
defined cost function) can be efficiently computed by max- ficients. For example, additional restrictions on the object boundary
flow/min-cut algorithms. This technique can be applied to smoothness or soft user-specified segmentation constraints cannot
N-dimensional images. Given user-specified object and back- be added readily. As for the intelligent paint, judging by the ex-
ground seed pixels, the rest of the pixels are labelled automat- amples supplied by the authors, the advantage of their method over
ically. the traditional ‘magic wand’ is in speed and number of user in-
teractions. As it appears from the algorithm description and pre-
GrabCut [Rother et al. 2004] extends graph-cut by introducing sented results, it is unlikely that intelligent paint would be capable
iterative segmentation scheme, that uses graph-cut for inter- of solving hard segmentation problems like in [Rother et al. 2004],
mediate steps. The user draws rectangle around the object [Boykov and Jolly 2001]. Precise object boundary estimation is
of interest - this gives the first approximation of the final ob- also questionable, because the finest segmentation level is obtained
ject/background labelling. Then, each iteration step gathers by initial tobogganing oversegmentation, which may not coincide
color statistics according to current segmentation, re-weights with actual object borders.
the image graph and applies graph-cut to compute new re- Speaking about medical images, the best performing method is
fined segmentation. After the iterations stop the segmentation random walker (judging by the provided examples). It leaves be-
results can be refined by specifying additional seeds, similar hind both watershed segmentation and region growing behind in
to original graph-cut. quality and robustness of segmentation. The quality of segmenta-
tion comparable to is graph cuts, but random walker is capable of
Medical images have their own unique properties. In many cases finding the solution for number of labels > 2. However, it is rather
they are graylevel and objects that should be segmented are very slow and its implementation is not an easy task. Also, method
different in their structure and appearance from the objects that are extension to achieve some special algorithm properties (i.e. con-
common in photo editing. Probably that is why much research ef- trollable boundary smoothness) is not straightforward. It should
fort was applied for developing specific and efficient segmentation be mentioned, that multi-labelling tasks can be solved by min-cut
methods for the medical images domain. Nevertheless, some seg- graph algorithms [Boykov et al. 2001], but no attempt to apply this
mentation techniques like Graph Cuts can successfully be applied multi-labelling method to interactive image segmentation is known
to medical images also. Some specific ‘medical’ segmentation tech- to us.
niques are reviewed further.

Marker-based watershed transformation uses watershed trans- 1.2 Proposed method

form, supported by user-specified markers (seeds) for seg- We take an intuitive user interaction scheme - user specifies certain
menting graylevel images [Moga and Gabbouj 1996]. Wa- image pixels (we will call them seed pixels) that belong to objects,
tershed transform treats the image as a surface with the relief that should be segmented from each other. The task is to assign
specified by the pixel brightness, or by absolute value of the labels to all other image pixels automatically, preferably achiev-
image gradient. The valleys of the resulting ‘landscape’ are ing the segmentation result the user is expecting to get. The task
filled with water, until it reaches the ‘mountains’. Markers statement and input data is similar to [Boykov and Jolly 2001] and
placed in the image specify the initial labels that should be [Grady and Funka-Lea 2004], however the segmentation instrument
segmented from each other. differs.
Our method uses cellular automaton for solving pixel labelling
Random walker [Grady and Funka-Lea 2004] given a small num- task. The method is iterative, giving feedback to the user while
ber of pixels with user-defined seed labels (labels number can the segmentation is computed. Proposed method allows (but not
be grater than 2), analytically determines the probability that requires) human input during labelling process, to provide dynamic
a random walker starting at each unlabelled pixel will first interaction and feedback between the user and the algorithm. This
reach one of the pre-labelled pixels. By assigning each pixel allows to correcting and guidance of the algorithm with user input
to the label for which the greatest probability is calculated, im- in the areas where the segmentation is difficult to compute, yet does
age segmentation is obtained. This method provides unique not require additional user effort where the segmentation is reliably
segmentation solution into connected segments, with some computed automatically.
robustness against ‘weak’ boundaries. A confidence rating
Important properties of our method, that we would like to outline
of each pixel’s membership in the segmentation is also esti-
are:
mated.
1. Capable of solving moderately hard segmentation tasks (see
Interactive region growing is a descendant of one of the classic
examples in this paper);
image segmentation techniques. Initially, the seed pixel in-
side the object of interest is specified, and then neighboring 2. Works with images of any dimension N ≥ 1;
pixels are iteratively added to the growing region, while they
conform to some region homogeneity criterium. The main 3. Performs multi-label image segmentation (the computation
problem is ‘leaking’ of the growing region through ‘weak’ time is not directly affected by the number of labels);
boundaries, however some treatment of this problem was pro-
posed [Heimann et al. 2004]. This method works with two 4. Is extensible, allowing construction of new families of seg-
labels only - object and background. mentation algorithms with specific properties;

The performance of described photo editing methods was eval- 5. Interactivity - as the segmentation is refined with each itera-
uated in [Rother et al. 2004] (except for the intelligent paint). The tion, user can observe the evolution and refine the segmenta-
authors have clearly shown, that methods based on graph cuts allow tion “on the fly”;
 6. The algorithm is simple in both understanding and implemen-
tation;

7. Using cellular automata allows fast parallel implementation;

The rest of the paper is organized follows. Section 2 describes

the proposed method in detail. Section 3 gives the results, section (a) (b) (c)
4 provides discussion and comparison with other methods. Finally
the conclusive remark is given in section 5. Figure 1: Segmentation of a color image. (a) Source image, (b)
user-specified seeds, (c) segmentation results

2 Method details
At iteration t + 1 cell labels l t+1
p and strengths θ pt+1 are updated
2.1 Basic method as follows:
Cellular automata (CA) were introduced by Ulam and von Neu-
Code 1 Automata evolution rule
mann [von Neumann 1966]. Since then they’ve been used to
model wide variety of dynamical systems in various application do- // For each cell...
mains, including image denoising and edge detection [Popovici and for ∀p ∈ P
Popovici 2002], [Hernandez and Herrmann 1996]. A cellular au- // Copy previous state
tomaton is generally an algorithm discrete in both space and time, l t+1
p = lp;
t

that operates on a lattice of sites p ∈ P ⊆ Z n (pixels or voxels in θ pt+1 = θ pt ;

image processing). // neighbors try to attack current cell
A (bi-directional, deterministic) cellular automaton is a triplet for ∀q ∈ N(p)
A = (S, N, δ ), where S is an non-empty state set, N is the neighbor- if g(kC~p − C ~q k2 ) · θqt > θ pt
hood system, and δ : SN → S is the local transition function (rule). l t+1
p = lq
t
This function defines the rule of calculating the cell’s state at t + 1 θ pt+1 = g(kC~p − C ~q k2 ) · θqt
time step, given the states of the neighborhood cells at previous
end if
time step t.
end for
Commonly used neighborhood systems N are the von Neumann
end for
and Moore neighborhoods:
• von Neumann neighborhood
n Where g is a monotonous decreasing function bounded to [0, 1],
N(p) = {q ∈ Z n : kp − qk1 := ∑ |pi − qi | = 1}; (1) we use a simple one:
i=1 x
g(x) = 1 − ; (4)
~ 2
max kCk
• Moore neighborhood
To supply an intuitive explanation to the pseudocode above we
N(p) = {q ∈ Z n : kp − qk∞ := max |pi − qi | = 1}; (2) can use biological metaphor. We can treat pixel labelling process as
i=1,n
growth and struggle for domination of K types of bacteria. The bac-
teria start to spread (grow) from the seed pixels and try to occupy
The cell state S p in our case is actually a triplet (l p , θ p , C~ p ) - the all the image. That is why we called the method ‘GrowCut’. The
label l p of the current cell, ‘strength’ of the current cell θ p , and cell rules of bacteria growth and competition are obvious - at each dis-
feature vector C~p , defined by the image. Without loss of generality crete time step, each cell tries to ‘attack’ its neighbors. The attack
we will assume θ p ∈ [0, 1]. force is defined by the attacker cell’s strength θq , and the distance
A digital image is a two-dimensional array of k × m pixels. An between attacker’s and defender’s feature vectors C ~q and C~ p . If the
unlabelled image may be then considered as a particular configura- attack force is greater than defender’s strength - the defending cell
tion state of a cellular automaton, where cellular space P is defined is ‘conquered’ and its label and strength are changed. The result
by the k × m array set by the image, and initial states for ∀p ∈ P are of these local competitions is that the strongest bacteria occupy the
set to: neighboring sites and gradually spread over the image.
l p = 0, θ p = 0, C~ p = RGB p ; (3) The calculation continues until automaton converges to stable
configuration, where cell states seize to change. The example of
where RGB p is the three dimensional vector of pixel’s p color in image segmentation is shown in figures 1 and 2. The method is
RGB space. The final goal of the segmentation is to assign each guaranteed to converge as the strength of each cell is a increasing
pixel one of the K possible labels. and bounded.
When user starts the segmentation by specifying the segmenta-
tion seeds, the seeded cells labels are set accordingly, while their
strength is set to the seed strength value (more about it in section
2.3). This sets the initial state of the cellular automaton.

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

Figure 2: Bacteria evolution steps. White - ‘object’ labelled bacte-

ria, black - ‘background’ labelled bacteria, gray - neutral territory.
(a) - step 1, (b) - step 10, (c) - step 25, (d) - step 40, (e) - step 65.
2.2 Adding controllable boundary smoothness One important difference from the methods based on graph cuts
is that seeds do not necessarily specify hard segmentation con-
The basic scheme is very simple and yet is able to achieve quality straints. In other words - user brush strokes need not to specify
segmentation (see figures 1, 4, 5). But in some images, the result- only the areas of firm foreground or firm background, but instead
ing segments boundary can be ragged, see figure 3 (a). It can be can adjust the pixels state continuously, making them ‘more fore-
acceptable, or even necessary when the task is to capture the small- ground’ or ‘a little more background’ for example. This gives more
est detail of the boundary (i.e. in medical applications), but this versatile control of the segmentation from the user part and makes
can be an unwanted artifact when editing a generic high-resolution the process tolerable to inaccurate paint strokes. This is especially
photo. To achieve smoother boundary, we propose an extension to helpful for segmenting accurate boundaries of objects like flowers
the automata in section 2.1. Local transition rule is modified with or plant leaves, like in figure 7.
two additional conditions. First: the cell, that has too many enemies The seed ‘strength’ is controlled by the initial strength θ 0 that is
around - enemiest (p) ≥ T1 is prohibited to attack its neighbors. Sec- set by the user’s brush strokes. Hard constraints specify the pix-
ond: the cell that has enemiest (p) ≥ T2 is forced to be occupied by els that should not change during the evolution, which is easily
the weakest of it’s enemies, no matter the cell’s strength. The ene- achieved by setting seed cells strength to 1. Soft constraints set
mies number is defined by: their initial strength values < 1 or just increase or decrease current
cell strength by some value.
enemiest (p) = max (
l=1,K
∑ 1) (5) Unlike graph cuts, adding seeds in already correctly segmented
q∈N(p),lqt 6=l tp area (i.e. near the border) can help to achieve better segmentation
in troubled areas. This is especially true when accurate border in
The thresholds values T1 , T2 control the boundary smoothness. blurry or camouflaged area is estimated.
Reasonable values for 2D images and Moore neighborhood range
from 9 (no smoothing) to 6. The examples of segmented border for
identical initial seeds and different threshold values are presented
in figure 3. 3 Results
We demonstrate our segmentation method in several examples in-
cluding photo editing and medical image segmentation. We show
original data and resulting segments generated by our technique, for
a given set of seeds.

3.1 Medical images

Figure 4 shows segmentations that have been obtained on several
(a) (b)
medical images from the DICOM image samples [Barre n. d.] -
user-specified seeds and resulting segmentation.
Figure 3: Different boundary smoothness restrictions. (a) T1 , T2 =
9, (b) T1 , T2 = 6.

2.3 User interaction

For the discussion on user interaction we will consider the case of
two labels - object and background. To our opinion, this is the most
often case of segmentation task in photo editing. However, this is
only for the sake of clarity, everything said here is true for the case
of labels K > 2 also.
The segmentation is started by specifying the initial seeds. This
is done by user’s strokes with ‘object’ and ‘background’ brushes
(see figure 1.b - red pixels correspond to ‘object’ brush strokes, blue
- to ‘background’ strokes). Each paint stroke of a defined brush sets
the initial labels and strengths of seed pixels.
After the initial seeds are set, the automata evolution starts. The Figure 4: Medical images segmentation results.
initial, incomplete user-labelling is often sufficient to allow the en-
tire segmentation to be completed automatically, but by no means
always. While the cell labels are being computed, the user can
observe the progress and interactively correct and guide the la- 3.2 Photo editing
belling process if necessary. User editing takes the form of brushing
pixels for adding new segmentation constraints. Each new paint In Figures 5 and 7 we show results of segmentation of several pho-
stroke changes the states of the underlying pixels and affects the tographs.
automaton evolution. This strategy allows to extract regions of in- Naturally, the hope is that the method can quickly identify the
terest from complex backgrounds using simple paint strokes with a right object, with little user interaction required. If the user needs
mouse. to specify too many seeds, this is not much better than a manual
Just as in [Rother et al. 2004] it is usually sufficient to brush, segmentation. The performance of an algorithm can be assessed by
roughly, just part of a wrongly labelled area. Note that correction the amount of efforts required from the user. Thus, the results of
can be done not only after the segmentation process is finished, but our segmentation are shown along with seeds that we entered by
in ‘the middle’ of the segmentation computation. the user.
 method by ourselves, but the results presented in the paper [Grady
and Funka-Lea 2004] and in the presentation poster are very close
to what we’ve been able to achieve on similar images.
Generally, methods based on graph cuts tend to produce
smoother (actually, shorter) boundary than our basic method. How-
ever our modified algorithm with controllable boundary smooth-
ness (see section 2.2) makes this difference less perceptible.
Several authors mention that segmentation methods, based on
minimum graph cuts tend to cut away small isolated image seg-
ments [Grady and Funka-Lea 2004], [Veksler 2000]. We performed
some investigation on how often this problem actually arises. In our
experience, this situation occurs only in cases when regional term
R(a) in eq. (1) in [Boykov and Jolly 2001] becomes insignificant
compared to boundary term B(A). One such example is shown in
figure 6, one can identify both smoother boundary of the graph cuts
segmentation, and isolated background segments on the right part
of the image. This can usually be avoided by tuning the relative
importance coefficient λ .

(a) (b) (c)

Figure 5: Generic photos segmentation results. Figure 6: Graph cuts and GrowCut segmentation difference for
same seeds. The Graph cuts produce smoother boundary, but has
problems with the background on the right image part, where two
4 Discussion isolated clusters instead of the whole region were segmented

As we have already emphasized in the introduction, our hope is It also should not be forgotten that both Random Walker and
to stir up the research community, motivating to search new ideas GrowCut are capable of solving multi-labelling tasks.
in the field of cellular automata and evolutionary computation and Random Walker boasts of creating segments connected to the
applying them to interactive image segmentation. initial seeds always. So as GrowCut does. Graph Cut potentially
We expect that results exceeding our current can be obtained. can (and sometimes does) create segments not connected to ini-
However, our current method can already compete with elegant tial seeds, if the regional term R(A) is stronger than boundary term
achievements of graph theory. In this section we will try to compare B(A). Same applies GrabCut. However, whether this an advantage
current top performing methods with ours and point out advantages or a drawback is probably the question of exact segmentation task.
and disadvantages of our scheme.
We take four methods - Graph Cuts, GrabCut, Random Walker
and GrowCut and compare them by several criteria: segmentation 4.2 Speed
quality, speed and convenience for the user. Accurately speaking,
The method speed is surely important. The reaction time to the
the methods differ seriously by the amount of information that they
user input should be small enough not to produce discomfort and
extract from the image. GrabCut uses most information - it com-
irritation, otherwise the method cannot be called interactive. An
putes the evolving color statistics of foreground and background
important measure is not only how fast the initial segmentation is
and takes into account color difference between neighboring pix-
achieved, but also how fast can the segmentation be recomputed,
els. Graph Cuts differs in using color statistics collected from the
when user adds more seeds to refine the initial segmentation.
user-specified seeds only, computed before the segmentation start.
The fastest are algorithms based on graph cuts. The initial seg-
Random Walker uses only intensity difference between neighboring
mentation computation time for Graph Cuts reported by the authors
pixels. Our current GrowCut variant also does not take advantage of
is ‘less than a second’ for ‘most 2D images (up to 512x512)’ on
object color statistics, however it can be easily extended to maintain
‘333MHz Pentium III’. This looks terrific, but judging from the
regions color statistics and use them in automaton evolution.
Intel processor manufacturer site, the frequency for Pentium III
processor family ranges from 450 MHz to 1.33 GHz. This sug-
4.1 Segmentation quality gests that actually running times are given for 1.33 GHz processor,
which is still a very good result. GrabCut processing time was given
Judging by our own experiments, Graph Cuts, GrabCut (we used for 450x300 image and equals to 0.9sec on 2.5GHz CPU with 512
our own implementation of both methods) and GrowCut are all able Mb RAM. Recomputing segmentation after new user-added seeds
to produce high quality segmentation in natural and medical im- is even faster than that. However, our own implementation of these
ages, so in this case mostly the amount user effort required should methods, based on Vladimir Kolmogorov’s publicly available code
be the measure of performance. However, there are some inter- [Boykov and Kolmogorov 2004] was not able to reach this speed.
esting features in these methods results, mentioned in next para- Random walker cannot boast of such speed, for the computation
graph. Concerning Random Walker, we’ve not been able to test the time for 256x256 image on dual Intel Xeon 2.4GHz processor with
1GB RAM is almost 5 (4.831) seconds. No indications that adding
more seeds will need less time to compute a refined segmentation
were found in the method description. But again, Random Walker
solves a harder task of multi-labelling.
Our current implementation goes somewhere in between. It
takes about 4 seconds on 256x256 image on 2.5GHz processor.
However, this is the total automaton evolution time (till complete
stop). Usually the desired segmentation result is obtained much
earlier (about 1.5-2 seconds) and the later automaton evolution does
not change the segmentation and can be stopped by the user, when
he is satisfied with the result. Another important issue is that our
methods allows user intervention at any iteration step, so adding
additional seeds throughout the computation does not increase the
whole working time. Again - our method solves the task of multi-
labelling.

4.3 User convenience

In our opinion, most important performance measure is the con-
venience of the image segmentation instrument, based on the algo-
rithm. Or, in other words, how much user effort is needed to achieve
satisfying segmentation result.
The results reported by Boykov et al. [Boykov and Jolly 2001]
say that it takes from 1 to 3-4 minutes to segment an image with
Graph Cuts, depending on the scene. Our own experiments have
shown that usually it takes about 1-2 minutes for the user, armed
with GrowCut to segment an image like shown in figures 4, 5, 7.

5 Conclusion and future work

In this paper, we have proposed a novel algorithm for interactive im-
age segmentation, based in cellular automaton. Experiments have
shown, that user effort required for segmenting generic photos and
medical images is rather modest, no harder than in state-of-the-
art graph-based methods like Graph Cuts, GrabCut and Random
Walker.
Proposed method combines the advantages, distributed between
the mentioned methods (multi-label segmentation, N-dimensional
images processing, speed high enough for interactive segmenta-
tion), and offers more - algorithm extensibility by varying the au-
tomaton evolution rule, more interactivity and user control of the
segmentation process.
Future work will concentrate on exploring theoretical properties
of the algorithm, research into exploiting color statistics during seg- Figure 7: More photo segmentation results.
mentation, and practical issues like faster implementation.

6 Acknowledgements B OYKOV, Y., V EKSLER , O., AND Z ABIH , R. 2001. Fast approximate
energy minimization via graph cuts. IEEE Trans. Pattern Anal. Mach.
We gratefully acknowledge valuable comments from the Graphicon Intell. 23, 11, 1222–1239.
reviewing committee members, Yuri Boykov and Olga Veksler. We
G RADY, L., AND F UNKA -L EA , G. 2004. Multi-label image segmentation
also would like to thank our lab’s head Yuri Bayakovsky for attract- for medical applications based on graph-theoretic electrical potentials.
ing our attention to the field of cellular automata some years ago. In ECCV Workshops CVAMIA and MMBIA, 230–245.

H EIMANN , T., T HORN , M., K UNERT, T., AND M EINZER , H.-P. 2004.
References New methods for leak detection and contour correction in seeded region
growing segmentation. In 20th ISPRS Congress, Istanbul 2004, Inter-
BARRE , S. Medical image samples. national Archives of Photogrammetry and Remote Sensing, vol. XXXV,
http://www.barre.nom.fr/medical/samples/. 317–322.
B OYKOV, Y., AND J OLLY, M.-P. 2001. Interactive graph cuts for optimal
boundary and region segmentation of objects in n-d images. In Proc. of H ERNANDEZ , G., AND H ERRMANN , H. J. 1996. Cellular automata for
the International Conference on Computer Vision, vol. 1, 105–112. elementary image enhancement. CVGIP: Graphical Model and Image
Processing 58, 1, 82–89.
B OYKOV, Y., AND KOLMOGOROV, V. 2004. An experimental comparison
of min-cut/max-flow algorithms for energy minimization in vision. IEEE M OGA , A., AND G ABBOUJ , M. 1996. A parallel marker based watershed
Trans. Pattern Anal. Mach. Intell. 26, 9, 1124–1137. transformation. In ICIP96, II: 137–140.
M ORTENSEN , E. N.,View
AND BARRETT, W. A. 1998. Interactive segmenta-
publication stats
tion with intelligent scissors. Graphical Models and Image Processing
60, 5, 349–384.
M ORTENSEN , E. N., AND BARRETT, W. A. 1999. Toboggan-based intel-
ligent scissors with a four-parameter edge model. In CVPR, 2452–2458.
P OPOVICI , A., AND P OPOVICI , D. 2002. Cellular automata in image pro-
cessing. In Fifteenth International Symposium on Mathematical Theory
of Networks and Systems.
R EESE , L. 1999. Intelligent Paint: Region-Based Interactive Image Seg-
mentation. Master’s thesis, Department of Computer Science, Brigham
Young University, Provo, UT.
ROTHER , C., KOLMOGOROV, V., AND B LAKE , A. 2004. Grabcut - in-
teractive foreground extraction using iterated graph cuts. Proc. ACM
Siggraph.
V EKSLER , O. 2000. Image segmentation by nested cuts. In Proc. IEEE
Conf. Computer Vision and Pattern Recognition, vol. 1, 339–344.
VON N EUMANN , J. 1966. Theory of Self-Reproducing Automata. Univer-
sity of Illinois Press. Ed. and Completed by A. Burks.

7 About the authors

Vladimir Veznevets is a research fellow at Graphics and Media
Laboratory of Moscow State Lomonosov University. He graduated
with honors Faculty of Computational Mathematics and Cyber-
netics of Moscow State University in 1999. He received his PhD
in 2002 in Moscow State University also. His research interests
include image processing, pattern recognition, computer vision and
adjacent fields. His email address is [email protected].

Vadim Konouchine is a third year student of Moscow State

Lomonosov University, Faculty of Computational Mathematics and
Cybernetics. His research interests lie in the fields of statistics and
image processing. His email address is [email protected].