I.J.

Image, Graphics and Signal Processing, 2012, 5, 1-13

Published Online June 2012 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2012.05.01

A Review on Graph Based Segmentation

K. Santle Camilus, V. K. Govindan
Department of Computer Science and Engineering, National Institute of Technology Calicut
Calicut, India.
E-mail: [email protected]

Abstract— Image segmentation plays a crucial role in boundary tracking, region growing from internal
effective understanding of digital images. Past few and external seeds, an image-graph cut, etc).
decades saw hundreds of research contributions in this Humans and object models, such as a probabilistic
field. However, the research on the existence of general atlas [9] or an active shape model [10], are better for
purpose segmentation algorithm that suits for variety of object recognition than computers, but the other way
applications is still very much active. Among the many around is true for object delineation. For instance,
approaches in performing image segmentation, graph humans can easily select a point inside the object, but
based approach is gaining popularity primarily due to its they have difficulties in manual tracing the same
ability in reflecting global image properties. This paper boundary several times. In this sense, the most effective
critically reviews existing important graph based approaches are interactive, because they combine the best
segmentation methods. The review is done based on the model for object recognition (the user expert) with the
classification of various segmentation algorithms within best model for object delineation (some good algorithm).
the framework of graph based approaches. The major This also makes important the combination of object
four categorizations we have employed for the purpose of models for recognition with delineation algorithms for
review are: graph cut based methods, interactive methods, automatic segmentation [11, 12, 13]. Note that, in this
minimum spanning tree based methods and pyramid scenario, graph-based methods present different graph
based methods. This review not only reveals the pros in representations, where the nodes may be pixels [3, 5, 14,
each method and category but also explores its 6, 7, 1, 2, 8], pixel vertices [15, 4], regions [16], or even
limitations. In addition, the review highlights the need user-drawn markers [17]. They also differ in the graph
for creating a database for benchmarking intensity based algorithm used to solve the problem: graph matching [17],
algorithms, and the need for further research in graph random walker [18], the min-cut/max-flow algorithm [19,
based segmentation for automated real time applications. 20, 21, 22, 8], Dijkstra's algorithm [3, 15, 4, 23, 14, 6, 7,
2], Kruskal's or Prim's algorithm [24, 25, 26, 27], etc.
Index Terms— Image segmentation, graph based method, Note also that, a same graph-search algorithm (e.g.,
boundary detection, graph partitioning algorithm, Image Dijkstra's algorithm) can be used for region-based and
analysis boundary-based segmentation, besides other image
operators [23]. Therefore, the best way to differ graph-
based segmentation methods is by taking into account the
I. INTRODUCTION graph model and algorithm that they use to solve the
Image segmentation can simply result an image segmentation problem. Besides, there are several other
partition composed by relevant regions, but the most approaches in the literature [28, 29, 30], which were not
fundamental challenge in segmentation is to precisely presented as graph-based approaches, but they use similar
define the spatial extent of some object, which may be concepts and can be more efficiently and/or effectively
represented by the union of multiple regions. In this sense, implemented as a graph-based approach [31, 32].
examples of possible strategies are: Image segmentation produces a set of homogeneous
(i) The image is divided into regions and then the regions of an image such that all pixels of a region are
object is composed by the union of some of these desired to be connected. The integration of all these
regions [1, 2]. regions constitutes the entire image. Each region has a set
(ii) The approximate location of the object/boundary of pixels and each pixel is characterized by its position
is found and its spatial extent is defined from that and feature vector. All pixels of a region are similar with
location [3, 4, 5, 6, 7, 8]. respect to a set of features. The basic principle of most of
In any case, the strategies are composed by the graph based segmentation methods is graph
(a) Object recognition tasks (e.g., regions that partitioning. Each method treats an image as a graph G in
compose the object, a point on its boundary, a line which vertices are composed of pixels. Each edge has a
inside it, the verification of a segmentation result, weight generally determined based on the vertices it
the matching between the image and an object relates. In graph theory sense, the above segmentation
model, etc). concept is similar to finding a set of sub-graphs
(b) Object delineation tasks (e.g., the image partition {SG1 , SG2 ,..... SGn } from the graph G such that for all
into regions, pixel classification, optimum k {1,2..., n} , i, j and i  j, vi , v j  SGk with walks

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
2 A Review on Graph Based Segmentation

between vi and vj. The compounding of all the vertices of the review and suggestions for future direction of
all the sub-graphs equals the complete set of vertices of research.
the graph. Every sub-graph comprises of a collection of
vertices with strong affinities among them. A pictorial
II. GRAPH CUT BASED METHODS
representation demonstrating this relation between image
segmentation and graph partitioning can be seen in Graph cuts started with the work of Greig [34]. The
Figure 1. graph G can be partitioned into two connected
components A and B such that A  B  V and A  B  
by omitting the edges linking these two components. The
degree of association between A and B can be inferred
from the total weight of the discarded edges, which is
simply called as a graph cut (a pictorial representation of
the graph cut is shown in Figure 2.
cut ( A, B )   w(u, v)
uA
(1)

vB

Figure 1. Association between image segmentation and graph

partitioning.
(a) Image (b) Graph (c) Graph partition (d) Segmented image

The graph partitioning to achieve image segmentation

is a challenging problem due to the following questions
[33]:
1. What is the precise criterion for a good partition?
2. How can such a partition be computed efficiently? Figure 2. A graph cut
Due to the subjective nature, defining a good
partitioning in segmentation still remains in debate. An optimal bi-partition minimizes this graph cut value
Assuming this is done, determining a criterion that [33]. By suitably and repeatedly partitioning the graph
optimally does this is challenging. Methods that use constructed from an image using the graph cut, different
graphs for image segmentation have been widely homogeneous regions could be obtained. In another way,
investigated within the fields of image processing and each vertex could be considered as a region. By utilising
image understanding. In these methods, segmentation graph cut values, which is a measure to show how much
problems by analogy are translated into graph based two neighboring regions are homogeneous, regions could
problems and that are solved as the graph partitioning be united repeatedly to form image partitions. Major
problem. These graph based segmentation methods might works in this category are the minimum cut algorithm
be grouped as (1) graph cut based methods, (2) introduced by Wu and Leahy [35], a normalized cut
interactive methods, (3) minimum spanning tree based algorithm by Shi and Malik [33, 36], a variant of
methods and (4) pyramid based methods. normalized cut suggested by Sharon et al. [37], a
The review is organized into eight sections. Section 1, polynomial time solution to the variant of normalized cut
being this introduction describing the graph based by Hochbaum [38], methods [39, 17, 40] incorporating
approach of segmentation problem. Sections 2 to 5 priors to improve performance of the normalized cut, the
respectively reviews exhaustively the four classes of average cut by Sarkar et al.[ 41], the min-max cut by
methods, namely, graph cut based methods, interactive Ding et al. [42], and the optimum cut by Li and Tian [43].
methods, minimum spanning tree based methods and The other important contributions in this category are the
pyramid based methods. Graph based methods not minimum mean cut [44] and ratio cut [45] by Wang et al.,
belonging to the above four categories are briefly two methods to determine the global minima of an
reviewed in Section 6. The need for creating a database energy function by Jermyn and Ishikawa [46], an active
of images and its ground truth for benchmarking an contours based method by Boykov et al. [47], a heuristic
algorithm in intensity based segmentation research is isoperimetric ratio algorithm by Grady and Schwartz [48],
highlighted in Section 7, and finally the paper is a branch-and-bound based technique by Lempitsky et al.
concluded in Section 8 highlighting the major outcome of [49], Watershed cuts by Cousty et al. [50], ratio regions
by Cox et al. [51], min-cut/max-flow algorithms[19-22],

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 3

and a few application focused methods [52-57]. These A variant of normalized cut was discussed by Sharon
methods are surveyed briefly in the following. et al. [37]. It is to minimize the ratio of similarity
Wu and Leahy [35] introduced minimum cut for image between the set of regions and its complement and the
segmentation in such a way that the smallest (k-1) cuts similarity within the set of regions. Recently, Hochbaum
among all possible cuts are selected and the [38] devised a polynomial time solution for this problem
corresponding edges are deleted to form k-subgraph which uses a graph cut procedure. A few methods have
partitions. This method favors the formation of very also been proposed to improve the performances of the
smaller regions, which results in over-segmentation. To normalized cut by incorporating priors. Yu and Shi [39]
overcome this problem, Shi and Malik [33, 36] proposed included a smoothed partial grouping constraint to the
a segmentation method based on normalized cut, the cut normalized cut. Eriksson et al. [17] admitted linear
cost function that is computed as a fraction of the total grouping constraints through a Lagrangian dual
edge connections to all the vertices in the graph. The formation. A constraint normalized cut, which adds prior
normalized cut, which is a partition criterion of a graph information in data with explicit linear constraints using
and reflects the global impression of an image, is given an iterative algorithm, was proposed by Xu et al. [40].
as The eigenvectors of graph Laplacian or their
cut ( A, B) cut ( A, B) deviations are normally used to partition a graph in
Ncut ( A, B)   (2) spectral graph partitioning. Sarkar et al. [41] presented a
assoc( A,V ) assoc( B,V )
spectral clustering technique based on average cut. This
Where assoc( A,V )   w(u, t ) is the total link measure is defined as the proposition of the total cut-link
uA,tV
weight normalized by the size of the partitions. Ding et al.
from vertices in A to all vertices in the graph. [42] presented a segmentation technique based on min-
assoc( B,V ) has similar definition. The authors showed max cut criterion. Li et al. [43] suggested optimum cut
that exact minimization of normalized cut is NP- criterion for partitioning a graph. The optimal solutions
complete. Also, they proposed an approximation to the min-max cut criterion and the optimal cut criterion
algorithm for the minimization of the cut cost function by are all NP-complete [43]. In [44], Wang and Siskind
solving a generalized eigenvalue problem. There are four presented a polynomial time algorithm for finding a
steps in the algorithm: (1) a graph is formed out of an minimum mean cut. An illustration is given in Figure 4.
image where vertices are pixels and weight of an edge is
a function of similarity between two neighboring pixels.
(2) The following eigenvalue system is solved for
eigenvectors with the smallest eigenvalues
( D  W ) x  Dx (3)
Where D is the diagonal matrix with d(i) =  w(i, j)
j
on its diagonal which is the total connection from vertex i
to all other vertices. W is the symmetrical matrix
with W (i, j )  wij . Also, x and  are the eignvector and
eignvalue respectively. (3) The graph is bi-partitioned
using the second smallest eignvector. (4) The above three
steps are recursively applied to the two partitions until
the normalized cut value is below a threshold. This
approximation algorithm is computationally expensive.
Also, the normalized cut tends to produce equally sized
regions which rarely occur in natural images. The results
obtained using this approximation algorithm is presented
in Figure 3. Figure 4. Segmentation of Lena and pepper images using the minimum
mean cut [44]

Later on Wang and Siskind [45] examined the

minimization of cut ratio cost function and discovered
that finding a minimum cut ratio in an arbitrary graph is
NP-hard. Hence, they generalized the polynomial time
algorithm for finding a minimum mean cut to obtain a
polynomial time algorithm for finding a cut that
minimizes the cut ratio cost function in connected planar
graphs. This solution repeatedly calls an inefficient non-
bipartite matching algorithm [38]. Figure 5 shows some
Figure 3. Results using the normalized cut [33, 36]. (a) Original image. results obtained on pepper and medical images.
(b) Segmented image

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
4 A Review on Graph Based Segmentation

Figure 6. Segmentation results produced by the isoperimetric algorithm

[48]

The min-cut/max-flow algorithms, which can be

Figure 5. The pepper and medical image segmentation produced by the
ratio region method [45] employed for energy functions’ minimization, have been
used to achieve image segmentation. Kolmogorov et al.
Jermyn and Ishikawa [46] suggested two methods to [59] provided necessary and sufficient conditions for
determine the global minima of an energy function for such energy functions. Geometric properties of regions
the identification of homogeneous regions in images. formed by graph cut methods are detailed in [47]. A few
This energy function is defined on the space of suggested min-cut/max-flow methods for image
boundaries in the image domain and can incorporate segmentation in the literature are: push-relabel methods
information both from the boundaries and interior of [19], the Dinic algorithm [20], the Boykov method [21]
regions. One limitation of this method is that it fails to and topological cuts [22]. The push relabel methods [19]
detect multiply-connected regions. Boykov et al. [47] maintain a preflow, a flow function. The methods run as
proposed a method by fusing both active contours and long as there is an active vertex in the graph. The push
graph cuts. This method demands the extraction of object operation increases the flow on a residual edge. A height
and background seeds which is hard to obtain for many function on vertices controls which residual edges can be
applications. It can be noted from [58] that determining pushed. The height function is altered with relabel
isoperimetric sets is a NP-hard problem. Grady and operations. The push and relabel operations guarantee
Schwartz [48] employed a heuristic algorithm for finding that the resulting flow is a maximum flow. A direct
a set with a low isoperimetric ratio in polynomial time, application of this method to image segmentation can be
which uses graph cuts, for image segmentation. An seen in the work of Ishikawa and Geiger [60]. The Dinic
instance is given in Figure 6. Lempitsky et al. [49] gave a algorithm [20] pushes flow on non-saturated paths from
global optimization framework for image segmentation the source to the sink until the maximum flow in the
using the graph cut and branch-and-bound techniques. graph is achieved. The Boykov method [21] works by
The worst case running time of this framework is higher. iteratively repeating the three stages: growing,
Cousty et al. [50] studied watersheds in edge-weighted augmentation and adoption. This method was found to be
graphs under the name watershed cuts. Graph cuts play a 2 to 5 times faster than the push-relabel methods and the
significant role in defining these watershed cuts. As the Dinic algorithm. The drawback of this method is that the
method has a drawback of forming numerous irrelevant augmenting paths found are not necessarily the shortest
smaller regions, it is required to employ a pre-processing augmenting paths. Zeng et al. [22] was the first to study
and\or a post-processing to obtain the desired outcome. about a min-cut/max-flow algorithm that incorporates a
Cox et al. [51] introduced ratio regions for image topological constraint: the topological cuts problem. It
segmentation. It is to minimize the ratio of the similarity was shown that any optimal solution to the topological
between the set of regions and its complement and the cuts problem is NP-hard. Hence, an approximation
number of regions within the set. A recent study by solution [22] was suggested to solve the topological cuts
Hochbaum [38] showed that this problem has a problem.
polynomial time solution using graph cuts. Chen et al. [52] used a modified graph cuts based
active contours as a fully automatic segmentation method
for RNAi fluorescence images. Yang et al. [53]
segmented densely packed cells in electron microscopic
images via graph cuts. Zhilan et al. [54] employed a
graph cut algorithm to segment arbitrary skin regions in
images. Hu et al. [55] extracted clothing by graph cuts.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 5

This method works based on several assumptions which [61] presented a discrete map data structure, which uses a
include cloth arms are similar in color to the torso, which region adjacency graph for merging operations, for
in turn contains dominant colors. These assumptions are interactive image segmentation. The discrete map permits
not always true as a cloth with sleeves in different colors a unified representation of both topology and geometry of
to the torso is quite common now. Han et al. [56] applied a segmented image.
a graph cut based segmentation method to extract white
mater, gray matter, and cerebral spinal fluid from brain
diffusion tensor image data. Camilus et al. [57] proposed
an approach to identify the pectoral muscle in
mammograms using a graph cut based merging method
and a Bezier curve algorithm. The result of this approach
is influenced by the order in which regions are merged in
the graph cut based merging method.
An unacceptable running time of a method could Figure 7. Illustration of the live wire [4, 15]. The two pre-defined points
prevent its usage in real time applications. Most of the in the boundary of the desired object are shown using arrows in each
methods in this category are computationally expensive case. (a) The detected boundary of the bone talus in an MR image of a
foot. (b) A detected vessel in an MR image of a wrist. (c) The detected
as they are proved to be NP complex and might not be internal boundary of the cortical part of a bone in the wrist
suitable for many real time usages.
The interactive graph cuts [8, 62] method is a popular
III. INTERACTIVE METHODS method of this category. This method exploits two
constraints for segmenting images: hard and soft.
A large variety of interactive segmentation methods Consider ρ contains pixels and N contains unordered
have been developed during the years. In general, none of pairs of neighbouring pixels. Let
them is superior to all the others. Also, some methods A  ( A1 , A2 , A3 ,... Ap ,....... A|| ) be a segmentation such
might be more suited for solving particular segmentation
problem better than the others. In short, the main steps of that Ap can either belong to the “background” or
an interactive graph based segmentation method are the “foreground”. The soft constraints are defined in the form
following: (1) Get the user preferences and (2) generate of a cost function E(A) that takes into account both
an optimal solution (if not, a sub-optimal solution) boundary and region properties of segments.
according to the user preferences and show it. In E ( A)  .R( A)  B( A) (4)
situations where automatic segmentation is difficult and Where
cannot guarantee correctness or reliability, these R ( A)   R p ( A p ) (5)
interactive methods are best opted. The interactive graph p

 B{ p,q} ( A p , Aq )
based segmentation methods take the advantage of
B ( A)  (6)
reliability under users’ control. Major research works in
{ p ,q}N
this category are live-wire methods [3, 4, 15], a discrete
map data structure by Braquelaire et al. [61], an And
interactive graph cuts method [8, 62], methods admitting 1 if A p  Aq
 ( A p , Aq )   (7 )
shape priors into the interactive graph cuts method [63- 0 otherwise
72], methods that improve running time of the interactive The coefficient Rp(.) and B{p,q} are region and
graph cuts method [73- 77] and methods which use the boundary terms that specifies the penalties for assigning
interactive graph cuts method in some applications [78, pixel p to "object" and "background", and a penalty for a
79]. Other research contributions in this category are the discontinuity between p and q respectively. The
synergistic arch weight estimation approach by Miranda coefficient  , which specifies the relative importance of
et al. [5], a Grabcut algorithm by Rother et al. [80], a
the region term against the boundary term, can have
random walker algorithm by Grady [18], an active
values greater than or equal to 0. The user has to tender
contours refining method by Xu et al. [81, 82], and a hard constraints based on his intension by marking some
progressive cut algorithm by Wang et al. [83, 84]. These
pixels as foreground/background as
are briefly reviewed in the following.
p  O, Ap  ' Object '
In the live-wire method [3, 15], the user is required to
specify some points in the desired object boundary. The q  B, Ap  ' Background '
shortest path (using Dijkstra's algorithm), which links all The idea is to compute the global minimum of the soft
these pre-defined points, is determined. This shortest path constraints among all segmentations A satisfying the hard
makes up the boundary which encloses the desired object. constraints. This is done using max-flow/min-cut analysis.
A faster version of live-wire was later introduced by From an image, a graph with two additional vertices, a
Falcao et al. [4] by devising a linear time graph searching “background” terminal and an “object” terminal, is built.
algorithm. Based on segmentation results of 492 medical The edge set comprises of two types of undirected edges:
images, it was observed that this faster version is about terminal links (t-links) and neighborhood links (n-links).
1.3 to 31 times faster than the live wire [4]. An example Each vertex has two t-links connecting to the terminals
for the live wire is provided in Figure 7. Braquelaire et al. and the link weights are defined by the regional term and

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
6 A Review on Graph Based Segmentation

hard constraints. Each pair of neighboring pixels is attempts have been made to improve its running time by
connected by an n-link and its weight is defined by the means of multilevel banded heuristics [73], flow
boundary term. The minimum cost cut of the graph forms recycling [74], capacity scaling [75], multi-scaling [76],
the partition between the object and the background. It or by planar graph cuts [77]. Weldeselassie and
can be noted that this method also belongs to the class- Hamarneh [78] extended the interactive graph cuts to
graph cut based methods. Results of this method are segment diffusion tensor MRI data by making use of
recorded in Figure 8. Apart from its obvious advantage of tensor calculus and tensor dissimilarity metrics. Malcolm
producing globally optimal solutions, the interactive et al. [79] employed the interactive graph cuts method to
graph cuts method gives rise to less satisfactory results segment multi-modal tensor valued images by taking into
on images dominated by desired weak contrast account the Riemannian geometry of the tensor space.
boundaries (meaning- these weak contrast boundaries are Miranda et al. [5] introduced an interactive method for
intended to be identified) and undesired strong contrast synergistic arch weight estimation, which considers both
boundaries. image attributes and object information. The arch weight
estimation finds one of its applications as a basic step in
the interactive graph cuts method.
The Grabcut algorithm [80] extracts the foreground of
an image, by utilizing a rectangular shaped user’s input
which roughly holds the foreground, by employing graph
cut iteratively. In the random walker algorithm [18],
some pixels should be pre-classified by the user. In the
subsequent steps, an unclassified pixel is assigned a
classified pixel label when a random walker has been
given the greatest probability on traversing first to the
Figure 8. Segmentation of photographs using the interactive graph cuts classified pixel from the unclassified pixel. An example
method [62]. The dark red regions and blue regions are user marked
hard constraints for foreground and background respectively, while the
of segmentation results obtained using this random
corresponding light colors represent segmentation using this method walker algorithm is given in Figure 9.

A few methods admitted shape priors into the

interactive graph cuts method to improve its accuracy.
Freedman and Zhang [63] embedded shape priors into
weights of edges by using a level set formation. Slabaugh
and Unal [64] incorporated an elliptical shape prior into
the interactive graph cuts method. A user is expected to
initialize the segmentation by marking a seed point in an
image which evolves later as an ellipse on employing this
approach. Zhang et al. [65] also integrated elliptical
shape priors into the interactive graph cuts method, but in
a different way from the earlier method, to segment
cervical lymph nodes on sonograms. Zhu-Jacquot et al.
[66] integrated geometric shape priors into the interactive
graph cuts method for kidney segmentation from MRI.
By making use of the shape information of the heart-
heart is a compact blob, Funka-Lea et al. [67] included
blob constraints to isolate the heart from CT scans. The
interactive graph cuts method biases towards shorter
boundaries. Das et al. [68], by adding compact shape
priors to the interactive graph cuts method, altered its
behavior to bias towards larger objects. The interactive
graph cuts method is weak in segmenting thin elongated Figure 9. Results produced by the random walker algorithm [18]. (a)
objects. To overcome this problem, Vicente et al. [69] Original image. (b) User defined seeds: foreground- green and
background- blue. (c) Probability at each pixel that a random walker
enforced connectivity priors to it. The work also imposes released from that pixel reaches the foreground seed. (d) Outlined
topology [22], specifically, 0-topology. It was shown that foreground (red) by the algorithm
the connectivity constraint optimization problems are all
NP-hard [69]. Veksler [70] introduced a star shape prior Xu et al. [81, 82] suggested a segmentation method to
into the graph cuts method. Liu et al. [71] inserted a refine active contours by iteratively deforming contours
regional and contour generic shape prior into the graph using graph cuts. This method also allows an interactive
cuts framework. Kumar et al. [72] formulated OBJ CUT, correction of the final boundary if it is not satisfactory.
which uses a loopy belief propagation algorithm to Examples of the segmentation found by the Graph cuts
incorporate shape priors into graph cuts. Aside improving based active contours are displayed in Figure 10. The
the accuracy of the interactive graph cuts method, a few drawback of this method is that it can segment a single

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 7

object but not multiple objects and also, the graph that is Zahn [85] used an approach based on Gestalt theory
used in this method must be constructed with appropriate [88] for detecting and describing clusters. From a given
pixel connectivity and edge weights. Wang et al. [83, 84] set of points, an MST is constructed first and inconsistent
proposed a progressive cut algorithm which explicitly edges in the MST are then deleted to obtain a collection
considers user’s intention into a graph cut framework for of connected components which in turn constitute
object cutout task. clusters. In the method of Xu et al. [86, 87], a tree
The interactive graph based methods demand user partitioning algorithm splits up the MST built from an
intervention which is not possible or desirable in many image into many sub-trees, which represent
applications. Also, necessary training to the user for homogeneous regions, such that each sub-tree should
interacting to a real time system that uses any of these have at least a specific number of vertices and any two
methods might be required to produce intended results. nearby sub-trees should feature significantly dissimilar
average gray levels. For noisy images, the method yields
low quality results due to the incorrect configuration of
the MST as an object might be contained in more than
one sub-tree due to noise.
Inspired by the work of Zahn, Felzenszwalb et al. [24,
25] presented an efficient graph based method for
segmenting images using MST and is widely used as it
runs in video rate in practice. The method works with the
assumption that edges between vertices in the same
segment should have relatively low weights than edges
between vertices in different segments. Initially each
vertex is considered as a segment. Then in a greedy way,
two segments are repeatedly selected to consider for
merging. Based on a comparison predicate, the decision
of merging the two segments is made. The comparison
predicate is defined based on the internal difference of a
segment and difference between two segments. The
internal difference of a segment(S) is the largest weight
in the minimum spanning tree of the segment, which is
given by
int( s )  max (we ) (8)
Figure 10. Graph cuts based active contours [82].The two images in the e  MST ( S , E )
top row show the initialization for graph cuts based active contours. The
two images in the bottom row show the corresponding segmentation The difference between two segments is calculated as
results using this method the minimum weight edge connecting the two segments.
When the difference between two segments is less than
or equal to the minimum of any of the internal difference
IV. MINIMUM SPANNING TREE BASED METHODS of the two segments, then the predicate allows the two
segments to be merged. The authors showed that the
A spanning tree of a connected undirected graph is a segmentation produced by this method is neither too
subgraph which links all the vertices of the graph and coarse nor too fine. Also, any attempt to change the
there should be exactly a single path between any two definition of the difference between two segments (say,
vertices. For the graph, many spanning trees could be median edge weight instead of minimum weight edge
found. However, the minimum spanning tree (MST) is a between the two segments) to make the method more
spanning tree whose total weight of edges is less than or robust lead to the solution NP-hard. As the method
equal to the total weight of edges of every other spanning merges two segments based on a single low weight edge
tree. A minimum spanning tree of a graph in which between them, there are possibilities that the results could
vertices are the pixels and edges represent the similarity considerably be affected by noise if no initial filtering of
between the vertices that it connects, constructed from an the image is done. A sample result obtained using this
image, represents the possible weakest connections. By method is shown in Figure 11. Fahad et al. [26] and
suitably removing the lowest weighted edges, different Zhang et al. [27] suggested some modifications to
partitions that have stronger inherent affinities could be improve its performance.
found. Major MST based approaches of segmentation are In practical scenarios, acquiring images without noise
the method based on Gestalt theory by Zahn [85], a tree is almost impossible due to the perplexed imaging
partitioning algorithm by Xu et al. [86, 87], an efficient environment. Usage of this category of methods in those
MST algorithm by Felzenszwalb et al. [24, 25] and noisy images without pre-processing that includes
methods [26, 27] improving the performance of the filtering might yield unacceptable segmentation as the
efficient MST algorithm. A short discussion of these MST based methods are very much susceptible to noise.
methods is as follow.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
8 A Review on Graph Based Segmentation

shift variance and inability to segment elongated objects.

A few segmentation methods have been proposed based
on irregular pyramids. Montanvert et al. [94] exploited a
hierarchy of region adjacency graph, which performs
stochastic decimation which in turn uses two binary state
variables and a random variable, to achieve segmentation.
This method yields different segmentation results
depending on different outcomes of the random variable
for the same input settings. To overcome this drawback,
Jolion and Montanvert [95] proposed an adaptive
F pyramid in which the random variable is replaced by an
igure 11. Results obtained for the efficient graph based method [24, 25]. interest variable in the decimation process. A localized
(a) Original image. An initial Gaussian filtering of the original image
was performed before segmentation by setting sigma=0.5. (b)
pyramid was proposed by Huart et al. [96] to achieve
Segmented image segmentation. The bounded irregular pyramid was
proposed by Marfil et al. [97] which combines features
from regular and irregular pyramids. Brun and Kropatsch
[98] combined a combinatorial pyramid and a union-find-
V. PYRAMID BASED METHODS based decimation algorithm to attain segmentation.
A review of pyramidal structures that are used for With the aid of empirical methods, Marfil et al. [16]
image segmentation can be found in [16]. The general showed that irregular pyramids yield better results than
framework involves the creation of a graph from the regular pyramids in segmenting objects. Though irregular
original image. From this base graph, a set of graphs pyramids resolve the problems posed by the regular
defined in multi-level of resolution, which can be pyramids, they possess unpredictable structures which
visualized as a pyramid, is built. The vertices and edges account for unbounded execution time for local
at level L+1 are formed from the reduction of vertices operations within each level.
and edges at level L using a reduction function. A level
of pyramid called as working level is chosen as the VI. OTHER METHODS
responsible level to yield segmentation. Based on the
working principle to build pyramids, they can be There are many other methods which do not belong to
classified into two categories: (1) regular pyramids and (2) any of the above categories of the graph based methods.
irregular pyramids Methods based on graph-searching principles [99-102]
are notable of this kind. A few methods were proposed
A. Regular pyramids for image segmentation based on graph shortest path
In regular pyramids, spatial relationships and the algorithms [103, 104, 23]. It was shown that shortest
reduction factor, which is defined as the ratio between the paths, random walker, and watershed were all actually
number of vertices at level L and the number of vertices the same algorithm with different norms [105]. The work
at level L+1, are constant and fixed, hence, the size and of Gomez et al. [106] utilized an iterative binary coloring
the layout of the structure of the pyramids are predicable. technique, which considers the changing behavior of
The first segmentation method based on regular pyramids adjacent pixels, for image segmentation. Granularity of
was proposed by Chen and Pavlidis [89]. Ping et al. [90] the output of this method mainly depends on number of
utilized a pyramid built using a Gaussian filter function iterations: a few iterations might develop a wrong
having adjustable filter scales. Burt et al. [91] proposed a segmentation, while too much iteration might yield
pyramid linking approach to attain image segmentation. useless results. Corso et al. [107] proposed an extended
But the accuracy of this method is sensitive to the right graph-shift algorithm, which is a hierarchical energy
selection of the working level. To overcome this minimization algorithm, for segmenting brain tumors and
drawback, a modified pyramid linking approach which multiple sclerosis lesions in MRI. Yuan et al. [108]
particularly uses two scaling rules was suggested by proposed a multi-modal segmentation method based on
Ziliani and Jensen [92]. Though this method attempted to region fusion and narrow band energy graph partitioning.
improve the pyramid linkage approach, a few drawbacks Ta et al. [109] utilized a framework of graph-based tools
remain in common to both these approaches. They are: for microscopic image segmentation.
elongated regions are not properly segmented, and the
structure of the pyramid varies even due to small VII. ISSUES IN PERFORMANCE EVALUATION
rotations, shifts and scales of the input image [93].
Each graph based segmentation technique addresses a
B. Irregular pyramids solution to some specific segmentation problems. Due to
In opposite to the regular ones, spatial relationships this, a technique might work well for a particular set of
and the reduction factors are not constants in irregular images but fail to segment satisfactorily for another set of
pyramids; hence, the size and the layout are not images. Authors tested their algorithms using limited
predictable. However, these types of pyramid solve the private set of images or by using standard test images like
problems associated with the regular pyramids such as Lena, photographer, baboon and so on and compare the

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 9

performance with other algorithms using qualitative maximizes the dissimilarities across boundaries is also a
evaluation or\and one or more quantitative criteria [110- study of interest to the research community.
112]. The main concern is whether the chosen dataset
represent all necessary test scenarios like different
illuminations, variations, conditions (say, day and night) REFERENCES
and contrasts. Given this, creating a database of images
[1] Noma, A., Graciano, A., Consularo, L., Cesar, M. R.,
and its ground truth for benchmarking an algorithm in
& Bloch, I. (2008). A new algorithm for interactive
intensity based segmentation would be an interesting area
of research for computer vision and image processing structural image segmentation. arXiv:0805.1854v2
[cs.CV].
community for two primary reasons: (1) to evaluate an
[2] Rocha, L.M., Cappabianco, F.A.M., & Falcao, A.X.
algorithm with different test images which represent all
possible test scenarios and (2) to compare different (2009). Data clustering as an optimum-path forest
algorithms in a common platform and rank them based problem with applications in image analysis.
on their performance. Without this dataset and ground International Journal of Imaging Systems and
truth, it is very difficult to evaluate and compare different Technology, 19(2), 50-68.
graph based segmentation methods. Though at this [3] Mortensen, E. (1998). Interactive segmentation with
intelligent scissors. Graphical Models and Image
moment a few benchmarking systems are publically
Processing, 60 (5), 349-384.
available, they are not very much suitable for evaluating
an algorithm in intensity based segmentation. For [4] Falcao, A. X., Udupa, J. K., & Miyazawa, F. K.
example, most of the test cases of the Berkeley (2000). An ultra-fast user-steered image
segmentation dataset and benchmarking [113] estimate segmentation paradigm: live wire on the fly. IEEE
the ability of an algorithm in identifying textures. Transaction on Medical Imaging, 19 (1), 55-62.
[5] De Miranda, P. A. V., Falcão, A. X., & Udupa, J. K.
(2010). Synergistic arc-weight estimation for
VIII. CONCLUSION interactive image segmentation using graphs.
Computer Vision and Image Understanding, 114
Recently, there has been increasing interest in using
(1), 85-99.
graph based methods as a powerful tool for segmenting
[6] Bai, X., & Sapiro G. (2007). Distance cut: interactive
images. This review has discussed some of the major
segmentation and matting of images and videos.
graph based methods and highlighted their strengths as
IEEE International Conference on Image
well as limitations. Some difficulties of these methods
Processing, San Antonio, Texas, 2, pp. 249-252.
have brought down their use in practical applications.
[7] Protiere, A., & Sapiro, G. (2007). Interactive image
The primary reason is the higher computational
segmentation via adaptive weighted distances.
complexity. The search for a vertex or an edge in a graph
IEEE Transactions on Image Processing, 16(4),
requires polynomial time. At the same time, search for a
1046-1057.
match of a scene model to an object model in the graph
[8] Boykov, Y. Y., & Jolly M. P. (2001). Interactive
has exponential complexity [114]. The current research in
graph cuts for optimal boundary & region
graph based methods orients towards producing
segmentation of objects in n-d images. Eighth IEEE
approximate solution (sub-optimal solution) for such
International Conference on Computer Vision, 1, pp.
graph matching problem to reduce processing time. Also,
105-112.
use of a priori information that include shape, topology
[9] Talairach, J., & Tournoux P. (1988). Co-planar
and appearance model of the category of images to be
stereotaxic atlas of the human brain. New York:
segmented is getting more popularity. Apart from the
Thieme Medical Inc.
class-wise disadvantages in general, in fact, many of the
[10] Cootes, T., Taylor, C., Cooper, D., & Graham, J.
graph based segmentation methods have their own
(1995). Active shape models-their training and
disadvantages. This survey demonstrates that most of the
application. Computer Vision and Image
methods fail to find use in automated real time
Understanding, 61(1), 38-59.
applications which normally work with no human
[11] Grau, V., Mewes, A.U.J., Alcaniz, M., Kikinis, R.,
involvement and demand acceptable running time with
& Warfield, S.K. (2004). Improved watershed
robust behavior of the methods. This demonstrates the
transform for medical image segmentation using
need for further research to refine the graph based
prior information. IEEE Transactions on Medical
segmentation techniques to be applied for automatic real
Imaging, 23(4), 447-458.
time scenarios. A few important future direction of
[12] Liu, J., & Udupa, J.K. 2009. Oriented active shape
research can be: (1) incorporation of the fuzzy set theory
models. IEEE Transactions on Medical Imaging,
into graph based frameworks to achieve enhanced
28(4), 571-584.
segmentation performances. (2) Use of multi-criteria to
[13] Miranda, P.A.V., Falcao, A.X., & Udupa, J.K.
partition a graph to achieve an efficient segmentation
(2009). Cloud bank: a multiple clouds model and its
solution is a promising direction of research. (3)
use in MR brain image segmentation. IEEE
Constructing a graph using feature sets rather than pixel
International Symposium on Biomedical Imaging,
level information and finding an optimum partition that
Boston, MA, pp. 506-509.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
10 A Review on Graph Based Segmentation

[14] Miranda, P.A.V., & Falcao, A.X. (2009). Links [28] Saha, P.K., & Udupa, J.K. (2001). Relative fuzzy
between image segmentation based on optimum- connectedness among multiple objects: theory,
path forest and minimum cut in graph. Journal of algorithms, and applications in image segmentation.
Mathematical Imaging and Vision, 35(2), 128-142. Computer Vision and Image Understanding, 82 (1),
[15] Falcao, A. X., Udupa, J. K., Samarasekera, S., 42-56.
Sharma, S., Hirsch, B. E., & Lotufo R. de. A. [29] Ciesielski, K.C., Udupa, J.K., Saha, P.K., & Zhuge
(1998). User-steered image segmentation paradigms: Y. (2007). Iterative relative fuzzy connectedness for
live wire and live lane. Graphical Models and multiple objects with multiple seeds. Computer
Image Processing, 60 (4), 233-260. Vision and Image Understanding, 107(3), 160-182.
[16] Marfil, R., Molina-Tanco, L., Bandera, A., [30] Vincent, L., & Soille, P. (1991). Watersheds in
Rodriguez, J. A., & Sandoval F. (2006). Pyramid digital spaces: an efficient algorithm based on
segmentation algorithms revisited. Pattern immersion simulations. IEEE Transaction on
Recognition, 39 (8), 1430-1451. Pattern Analysis and Machine Intelligence, 13(6),
[17] Eriksson, A. P., Olsson, C., & Kahl, F. (2007). 583-598.
Normalized cuts revisited: a reformulation for [31] Ciesielski, K.C., Udupa, J.K., Falcao, A.X., &
segmentation with linear grouping constraints. Miranda, P.A.V. (2011). Fuzzy connectedness and
International Conference on Computer Vision, pp. graph cut image segmentation - similarities and
1-8. differences. Proceedings of SPIE on Medical
[18] Grady, L. (2006). Random walks for image Imaging, To appear.
segmentation. IEEE Transactions on Pattern [32] Audigier, R., & Lotufo, R.A. (2007). Watershed by
Analysis and Machine Intelligence, 28 (11), 1768- image foresting transform, tie-zone, and theoretical
1783. relationship with other watershed definitions.
[19] Goldberg, A. V., & Tarjan R. E. (1988). A new Eighth International Symposium on Mathematical
approach to the maximum-flow problem. Journal of Morphology and its Applications to Signal and
the ACM, 35 (4), 921–940. Image Processing, Rio de Janeiro, Brazil, pp. 277-
[20] Dinic, E. A. (1980). Algorithm for solution of a 288.
problem of maximum flow in networks with power [33] Shi, J., & Malik, J. (2000). Normalized cuts and
estimation. Soviet Mathematics- Doklady, 11, image segmentation. IEEE Transactions on Pattern
1277-1280. Analysis and Machine Intelligence, 22 (8), 888-905.
[21] Boykov, Y., & Kolmogorov, V. (2004). An [34] Greig, D.M., Porteous, B.T., & Seheult A.H. (1989).
experimental comparison of min-cut/max- flow Exact maximum a posteriori estimation for binary
algorithms for energy minimization in vision. IEEE images. Journal of the Royal Statistical Society
Transactions on Pattern Analysis and Machine Series B, 51, 271-279.
Intelligence, 26 (9), 1124-1137. [35] Wu, Z., & Leahy, R. (2002). An optimal graph
[22] Zeng, Y., Samaras, D., Chen W., & Peng, Q. (2008). theoretic approach to data clustering: theory and its
Topology cuts: a novel min-cut/max-flow algorithm application to image segmentation. IEEE
for topology preserving segmentation in n–d images. Transactions on Pattern Analysis and Machine
Computer Vision and Image Understanding, 112 Intelligence, 15 (11), 1101-1113.
(1), pp. 81-90. [36] Shi, J., & Malik, J. (1997). Normalized cuts and
[23] Falcao, A. X., Stolfi, J., & Alencar, De. (2004). The image segmentation. IEEE Computer Society
image foresting transform: theory, algorithms, and Conference on Computer Vision and Pattern
applications. IEEE Transactions on Pattern Recognition, pp. 731-737.
Analysis and Machine Intelligence, 26 (1), pp. 19- [37] Sharon, E., Galun, M., Sharon, D., Basri, R., &
29. Brandt, A. (2006). Hierarchy and adaptivity in
[24] Felzenszwalb, P., & Huttenlocher, D. (2004). segmenting visual scenes. Nature, 442 (7104), 810-
Efficient graph-based image segmentation. 813.
International Journal of Computer Vision, 59 (2), [38] Hochbaum, D. S. (2010). Polynomial time
167-181. algorithms for ratio regions and a variant of
[25] Felzenszwalb, P.F., & Huttenlocher, D.P. (1998). normalized cut. IEEE Transactions on Pattern
Image segmentation using local variation. IEEE Analysis and Machine Intelligence, 32 (5), 889-898.
Computer Society Conference on Computer Vision [39] Yu, S. X., & Shi, J. (2004). Segmentation given
and Pattern Recognition, pp. 98-104. partial grouping constraints. IEEE Transactions on
[26] Fahad, A., & Morris, T. (2006). A faster graph- Pattern Analysis and Machine Intelligence, 26 (2),
based segmentation algorithm with statistical region 173-183.
merge. Lecture Notes in Computer Science, pp. [40] Xu, L., Li, W., & Schuurmans, D. (2009). Fast
286-293. normalized cut with linear constraints. IEEE
[27] Zhang, M., & Alhajj, R. (2006). Improving the Conference on Computer Vision and Pattern
graph-based image segmentation method. Recognition, pp. 2866-2873.
Eighteenth IEEE International Conference on Tools [41] Sarkar, S., & Soundararajan, P. (2000). Supervised
with Artificial Intelligence, pp. 617 - 624. learning of large perceptual organization: graph

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 11

spectral partitioning and learning automata. IEEE [56] Han, D., Singh, V., Lee, J. E., Zakszewski, E.,
Transactions on Pattern Analysis and Machine Adluru, N., Oakes, T. R., & Alexander, A. (2009).
Intelligence, 22 (5), 504-525. An experimental evaluation of diffusion tensor
[42] Ding, C. H. Q., He, X., Zha, H., Gu, M., & Simon, image segmentation using graph-cuts. Conference
H. D. (2001). A min-max cut algorithm for graph of the IEEE Engineering in Medicine and Biology
partitioning and data clustering. International Society, 5653-5656.
Conference on Data Mining, pp. 107-114. [57] Camilus, K. S., Govindan, V. K., & Sathidevi, P. S.
[43] Li, X., & Tian, Z. (2007). Optimum cut-based (2010). Computer aided identification of the
clustering. Signal Processing, 87(11), 2491-2502. pectoral muscle in digitized mammograms. Journal
[44] Wang, S., & Siskind, J. M. (2001). Image of Digital Imaging, 23 (5), 562-580.
segmentation with minimum mean cut. Eighth [58] Mohar, B. (1989). Isoperimetric numbers of graphs.
IEEE International Conference on Computer Vision, Journal of Combinatorial Theory Series B, 47 (3),
1, pp. 517-524. 274-291.
[45] Wang, S., & Siskind, J. M. (2003). Image [59] Kolmogorov, V., & Zabin, R. (2004). What energy
segmentation with ratio cut. IEEE Transactions on functions can be minimized via graph cuts? IEEE
Pattern Analysis and Machine Intelligence, 25 (6), Transactions on Pattern Analysis and Machine
675-690. Intelligence, 26 (2), 147-159.
[46] Jermyn, I., & Ishikawa, H. (2001). Globally optimal [60] Ishikawa, H., & Geiger, D. (1998). Segmentation by
regions and boundaries as minimum ratio weight grouping junctions. IEEE Computer Society
cycles. IEEE Transactions on Pattern Analysis and Conference on Computer Vision and Pattern
Machine Intelligence, 23 (10), 1075-1088. Recognition, pp. 125-131.
[47] Boykov, Y., & Kolmogorov, V. (2003). Computing [61] Braquelaire, J. P., & Brun, L. (1998). Image
geodesics and minimal surfaces via graph cuts. segmentation with topological maps and inter-pixel
Ninth IEEE International Conference on Computer representation. Journal of Visual Communication
Vision, 1, pp. 26-33. and Image Representation, 9(1), 62-79.
[48] Grady, L., & Schwartz, E. L. (2006). Isoperimetric [62] Boykov, Y., & Funka-Lea, G. (2006). Graph cuts
graph partitioning for image segmentation. IEEE and efficient n-d image segmentation. International
Transactions on Pattern Analysis and Machine Journal of Computer Vision, 70 (2), 109-131.
Intelligence, 28 (3), 469-475. [63] Freedman, D., & Zhang, T. (2005). Interactive graph
[49] Lempitsky, V., Blake, A., & Rother, C. (2008) . cut based segmentation with shape priors, IEEE
Image segmentation by branch-and-mincut. Computer Society Conference on Computer Vision
European Conference on Computer Vision, pp. 15- and Pattern Recognition, 1, pp.755-762.
29. [64] Slabaugh, G., & Unal, G. (2005). Graph cuts
[50] Cousty, J., Bertrand, G., Najman, L., and Couprie, segmentation using an elliptical shape prior, IEEE
M. (2009). Watershed cuts: minimum spanning International Conference on Image Processing, 2,
forests and the drop of water principle. IEEE pp. 1222-1225.
Transactions on Pattern Analysis and Machine [65] Zhang, J., Wang, Y., & Shi, X. (2009). An improved
Intelligence, 31 (8), 1362-1374. graph cut segmentation method for cervical lymph
[51] Cox, I. J., Rao, S. B., & Zhong, Y. (1996). Ratio nodes on sonograms and its relationship with node's
regions: a technique for image segmentation. shape assessment. Computerized Medical Imaging
International Conference on Pattern Recognition, 2, and Graphics, 33, 602-607.
pp. 557-564. [66] Zhu-Jacquot, J., & Zabih, R. (2007). Graph cuts
[52] Chen, C., Li, H., & Zhou, X. (2007). Automated segmentation with statistical shape priors for
segmentation of drosophila RNAi fluorescence medical images, Third International IEEE
cellular images using graph cuts. Lecture Notes in Conference on Signal-Image Technologies and
Computer Science, pp. 116-125. Internet-Based System, pp.631-635.
[53] Yang, H., & Choe, Y. (2009). Cell tracking and [67] Funka-Lea, G., Boykov, Y., Florin, C., Jolly, M. P.,
segmentation in electron microscopy images using Moreau-gobard, R., Ramaraj, R., & Rinck, D.
graph cuts, Sixth IEEE International Conference (2006). Automatic heart isolation for CT coronary
on Symposium on Biomedical Imaging, pp. 306- visualization using graph-cuts. IEEE International
309. Symposium on Biomedical Imaging, pp. 614-617.
[54] Zhilan, H., Guijin, W., Xinggang, L., & Hong, Y. [68] Das, P., & Veksler, O. (2006). Semiautomatic
(2009). Skin segmentation based on graph cuts. segmentation with compact shape prior. 3rd
Tsinghua Science and Technology, 14 (4), 478-486. Canadian Conference on Computer and Robot
[55] Hu, Z., Yan, H., & Lin, X. (2008). Clothing Vision, pp. 28-36.
segmentation using foreground and background [69] Vicente, S., Kolmogorov, V., & Rother, C. (2008).
estimation based on the constrained delaunay Graph cut based image segmentation with
triangulation. Pattern Recognition, 41 (5), 1581- connectivity priors. IEEE Conference on Computer
1592. Vision and Pattern Recognition, pp.1-8.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
12 A Review on Graph Based Segmentation

[70] Veksler, O. (2008). Star shape prior for graph-cut [86] Xu, Y., Olman, V., & Uberbacher, E. C. (1996). A
image segmentation, 10th European Conference on segmentation algorithm for noisy images. IEEE
Computer Vision, pp. 454-467. International Joint Symposia on Intelligence and
[71] Liu, C., Li, F., Zhang, Y., & Gu, H. (2009). Systems, pp. 220-226.
Interactive image segmentation based on hierarchical [87] Xu, Y., & Uberbacher, E. C. (1997). 2D image
graph-cut optimization with generic shape prior. segmentation using minimum spanning trees. Image
Lecture Notes in Computer Science, pp. 201-210. and Vision Computing, 15 (1), pp. 47-57.
[72] Kumar, M. P., Torr, P. H. S., & Zisserman, A. [88] Wertheimer, M. (1938). Laws of organization in
(2005). OBJ CUT. IEEE Computer Society perceptual forms- A source book of Gestalt
Conference on Computer Vision and Pattern psychology. London: Routledge & Kegan Paul.
Recognition, 1, pp. 18-25. [89] Chen, P.C., & Pavlidis, T. (1980). Image
[73] Lombaert, H., Sun, Y., Grady, L., and Xu, C. (2005). segmentation as an estimation problem, Computer
A multilevel banded graph cuts method for fast Graphics and Image Processing, 12 (2), 153-172.
image segmentation, Tenth IEEE International [90] Ping, Y., Runsheng, W., & Diannong, L. (1996). A
Conference on Computer Vision, 1, pp. 259-265. new image segmentation approach based on linked
[74] Kohli, P., & Torr, P. H. S. (2007). Dynamic graph pyramid. Proceedings of International Conference
cuts for efficient inference in Markov random fields. on Signal Processing, pp. 1118–1121.
IEEE Transactions on Pattern Analysis and [91] Burt, P., Hong, T., & Rosenfeld, A. (1981).
Machine Intelligence, 29 (12), 2079-2088. Segmentation and estimation of image region
[75] Juan, O., & Boykov,Y. (2007). Capacity scaling for properties through cooperative hierarchical
graph cuts in vision. IEEE 11th International computation. IEEE Transaction on Systems, Man
Conference on Computer Vision, pp.1-8. and Cybernetics, 11 (12), 802–809.
[76] Delong, A., & Boykov, Y. (2008). A scalable [92] Ziliani, F., & Jensen, B. (1998). Unsupervised
graph-cut algorithm for N-D grids. IEEE image segmentation using the modified pyramidal
Conference on Computer Vision and Pattern linking approach. International Conference on
Recognition, pp.1-8. Image Processing, 3, pp. 303-307.
[77] Schmidt, F. R., Toppe, E., & Cremers, D. (2009). [93] Bister, M., Cornelis, J., Rosenfeld, A. (1990). A
Efficient planar graph cuts with applications in critical view of pyramid segmentation algorithms.
computer vision. IEEE Conference on Computer Pattern Recognition Letters, 11(9), 605- 617.
Vision and Pattern Recognition, pp. 351-356. [94] Montanvert, A., Meer, P., & Rosenfeld, A. (1991).
[78] Weldeselassie, Y. T., & Hamarneh, G. (2007). DT- Hierarchical image analysis using irregular
MRI segmentation using graph cuts. Proceedings of tessellations, IEEE Transaction on Pattern Analysis
SPIE, pp. 1-9. Machine Intelligence, 13(4), 307-316.
[79] Malcolm, J., Rathi, Y., and Tannenbaum, A. 2007. [95] Jolion, J. M., & Montanvert, A. (1992). The adaptive
A graph cut approach to image segmentation in pyramid, a framework for 2D image analysis,
tensor space, Proceedings of Workshop Component Computer Vision, Graphics, and Image Processing,
Analysis Methods, pp. 18-25. 55(3), 339-348.
[80] Rother, C., Kolmogorov, V., & Blake, A. (2004). [96] Huart, J., & Bertolino, P. (2005). Similarity-based
Grabcut: interactive foreground extraction using and perception-based image segmentation. IEEE
iterated graph cuts. ACM Transactions on International Conference on Image Processing, pp.
Graphics, 23, 309-314. 1148-1151.
[81] Xu, N., Bansal, R., & Ahuja, N. (2003). Object [97] Marfil, R., Rodriguez, J. A., Bandera, A., &
segmentation using graph cuts based active Sandoval, F. (2004). Bounded irregular pyramid: a
contours. Computer Vision and Pattern Recognition new structure for colour image segmentation.
Conference, 2, 46-53. Pattern Recognition, 37 (3), 623-626.
[82] Xu, N., Ahuja, N., & Bansal, R. (2007). Object [98] Brun, L., & Kropatsch, W. G. (2003). Construction
segmentation using graph cuts based active of combinatorial pyramids. Lecture Notes in
contours. Computer Vision and Image Computer Science, pp. 1-12.
Understanding, 107 (3), 210-224. [99] Montanari, U. (1971). On the optimal detection of
[83] Yang, Q., Tang, X., Wang, C., Ye, Z., & Chen, M. curves in noisy pictures, Communications of
(2007). Progressive cut: an image cutout algorithm ACM, 14(5), pp. 335-345.
that models user intentions. IEEE Multimedia, 14 [100]Martelli, A. (1972). Edge detection using heuristic
(3), 56-66. search methods. Computer Graphics and Image
[84] Wang, C., Yang, Q., Chen, M., Tang, X., & Ye, Z. Processing, 1(2), pp. 169-182.
(2006). Progressive cut, Proceedings of the 14th [101]Martelli, A. (1976). An application of heuristic
Annual ACM international Conference on search methods to edge and contour detection,
Multimedia, pp. 251-260. Communications of the ACM, 19(2), 73-83.
[85] Zahn, C. T. (1971). Graph-theoretical methods for [102]Pope, D. L., Parker, D. L., Clayton, P. D., &
detecting and describing gestalt clusters. IEEE Gustafson, D. E. (1985). Left ventricular border
Transactions on Computers, C-20 (1), 68-86.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13
 A Review on Graph Based Segmentation 13

recognition using a dynamic search algorithm. Dr. K. Santle Camilus is currently

Radiology, 155 (2), 513-518. working as Technical lead at
[103]Sonka, M., Winniford, M. D., & Collins, S. M. Samsung India Software Operations,
(1995). Robust simultaneous detection of coronary Bangalore, India. He received his
borders in complex images. IEEE Transactions on Bachelor’s degree from Madras
Medical Imaging, 14 (1), pp. 151-161. University in information
[104]Udupa, J. K., Saha, P. K., & Lotufo, R. A. (2002). technology in the year 2003 and
Relative fuzzy connectedness and object definition: Master’s degrees in Computer
theory, algorithms, and applications in image Science and Engineering from Manonmaniam
segmentation. IEEE Transactions on Pattern Sundaranar University in the year 2005. He received his
Analysis Machine Intelligence, 24 (11), 1485-1500. PhD degree in medical image analysis from National
[105]Couprie, C., Grady, L., Najman, L., & Talbot, H. Institute of Technology, Calicut, India in the year 2011.
(2009). Power watersheds: a new image He has over 2 years of industrial experience and 1 year of
segmentation framework extending graph cuts, teaching experience. His research areas of interest
random walker and optimal spanning forest. include image processing, pattern recognition and
International Conference on Computer Vision, pp. medical image processing. He has over 12 research
731-738. publications in various international journals and
[106]Gomez, D., Montero, J., Yáñez, J., & Poidomani, C. conferences. He has reviewed papers for many
(2007). A graph coloring approach for image conferences and journals including Springer and Elsevier.
segmentation. Omega, 35 (2), 173-183. His biography appeared in the 29th edition of Who's Who
[107]Corso, J. J., Yuille, A., Sicotte, N. L., & Toga, A. in the World.
(2007). Detection and segmentation of pathological
structures by the extended graph-shifts algorithm, Dr. V. K. Govindan is currently
International Conference on Medical Image serving as Professor of Computer
Computing and Computer-Assisted Intervention, pp. Science and Engineering
985-993. Department and Dean Academic,
[108]Yuan, X., Situ, N., & Zouridakis, G. A. (2009). National Institute of Technology,
Narrow band graph partitioning method for skin Calicut, India. He received
lesion segmentation. Pattern Recognition, 42 (6), Bachelor’s and Master’s degrees in
1017-1028. Electrical Engineering from National Institute of
[109]Ta, V., Lézoray, O., Elmoataz, A., & Schüpp, S. Technology (the erstwhile Regional Engineering
(2009). Graph-based tools for microscopic cellular College), Calicut in the year 1975 and 1978, respectively.
image segmentation. Pattern Recognition, 42 (6), He was awarded PhD in Character Recognition from the
1113-1125. Indian Institute of Science, Bangalore, in 1989. He has
[110]Zhang, Y. (1996) . A survey on evaluation methods over 32 years of teaching experience in the capacity of
for image segmentation. Pattern Recognition, 29 (8), Lecturer (1979-87), Asst. professor (1987-98) and
1335-1346. Professor (1998 onwards). He was Head of the
[111]Unnikrishnan, R., Pantofaru, C., & Hebert, M. Department of Computer Science and Engineering during
(2007) .Toward objective evaluation of image January 2000 to August 2005. His research areas of
segmentation algorithms. IEEE Transactions on interest include medical imaging, agent technology,
Pattern Analysis and Machine Intelligence, 29 (6), biometrics based authentication, data compression, and
929-944. distributed computing. He has over 85 research
[112]Zhang, H., Fritts, J., & Goldman, S. (2008). Image publications in various international journals and
segmentation evaluation: A survey of unsupervised conferences, and authored several books on Operating
methods, Computer Vision and Image systems and Computer basics. He has reviewed papers
Understanding, 110 (2), 260-280. for many conferences and journals including IEEE
[113]Martin, D., Fowlkes, C., Tal, D., & Malik, J. (2001). Transactions and evaluated several PhD theses.
A database of human segmented natural images and
its application to evaluating segmentation
algorithms and measuring ecological statistics.
International Conference on Computer Vision, 2, pp.
416-423.
[114]Sanfeliu, A., Alquezar, R., Andrade, J., Climent, J.,
Serratosa, F. & Verges, J. (2002). Graph-based
representations and techniques for image processing
and image analysis. Pattern Recognition, 35 (3),
639-650.

Copyright © 2012 MECS I.J. Image, Graphics and Signal Processing, 2012, 5, 1-13