An interactive image segmentation using multiple user inputªs

IJRET: International Journal of Research in Engineering and Technology ISSN: 2319-1163
__________________________________________________________________________________________
Volume: 02 Issue: 03 | Mar-2013, Available @ http://www.ijret.org 318
AN INTERACTIVE IMAGE SEGMENTATION USING MULTIPLE USER
INPUT’S
Swathika P1
, Pradhiba Selvarani M2
, Arulmary S3
, Krishnalakshmi B4
1, 2, 3, 4
Assistant Professor, CSE, Kamaraj College Of Engineering and Technology College, Tamilnadu, Virudhunagar,
Swathika.me@gmail.com, mpradhiba@gmail.com, arulmarys@gmail.com, krishna11.kcet@gmail.com
Abstract
In this paper, we consider the Interactive image segmentation with multiple user inputs. The proposed system is the use of multiple
intuitive user inputs to better reflect the user’s intention. The use of multiple types of intuitive inputs provides the user’s intention
under different scenario. The proposed method is developed as a combined segmentation and editing tool. It incorporates a simple
user interface and a fast and reliable segmentation based on 1D segment matching. The user is required to click just a few "control
points" on the desired object border, and let the algorithm complete the rest. The user can then edit the result by adding, removing
and moving control points, where each interaction follows by an automatic, real-time segmentation by the algorithm. Interactive
image segmentation involves a proposed algorithm, Constrained Random walks algorithm. The Constrained Random Walks algorithm
facilitates the use of three types of user inputs. 1. Foreground and Background seed input 2. Soft Constraint input 3. Hard Constraint
input. The effectiveness of the proposed method is validated by experimental results. The proposed algorithm is algorithmically
simple, efficient and less time consuming.
Keywords: Interactive image segmentation, Interactive image segmentation, digital image editing, multiple user inputs,
random walks algorithm.
---------------------------------------------------------------------***-------------------------------------------------------------------------
1. INRODUCTION
Interactive image segmentation, an important branch of image
segmentation aims to make use of a small amount of user
interaction to extract the foreground object out from the
background. The main challenge of interactive image
segmentation systems is to get as much information as possible
from limited user inputs. Interactive image segmentation
involves minimal user interaction to incorporate user intention
into the segmentation process. The users only need to roughly
indicate the location and region of the object and background by
using strokes, which are called markers. The user has to
initialize the foreground and background seeds to separate the
foreground and background objects.
For good interactive image segmentation, there are two basic
requirements: First, given a certain user input, the constrained
random walks algorithm should produce intuitive segmentation
that reflects the user intent. Second, the algorithm must be
efficient so that it can produce better segmented results.
Interactive segmentation minimizes user interaction, and also
allows users to draw strokes matting on the boundary of the
selected object. Completely automatic segmentation never
seems to be perfect and so interactive tools are becoming more
and more popular. This means that if the results obtained by the
first segmentation are not satisfactory, the user can refine the
selection. The process of extracting the foreground from the
background of an image is known as segmentation. There are
many tools that already exist which perform this function.
These tools either use texture (color) information or edge
(contrast) information to perform the segmentation. Magic
wand (color) and Intelligent scissors (contrast) are two good
examples. Early interactive image segmentation algorithms
utilize either regional properties such as Adobe‟s magic wand or
boundary properties such as active contour [7] and intelligent
scissors [10], [11]. The proposed Interactive segmentation
method introduces an innovative approach of using both color
and contrast information in order to do the segmentation.
The purpose of interactive segmentation is to split an image into
two parts – the background and foreground. In order to do this
the user has to give „clues‟ to the algorithm indicating which
regions of the image are part of the foreground and which are
part of the background. This can be done by using „seeds‟. A
pixel can either be labeled as definitely background or definitely
foreground. This is a means of hard segmentation. In other
words that pixel must become part of the foreground or
background. These pixels then become seeds. Seeds allow the
algorithm to automatically calculate the rest of the foreground
or background while making sure that the hard constraints are
met.
In this paper, the foreground image has to be segmented
accurately in still images using constrained random walks
algorithm. In Constrained random walks algorithm, three types

__________________________________________________________________________________________
of constraints are given as an input to the segmentation process.
The user has to draw the strokes manually, in the foreground
and in Background region. Three cases arise for providing a
input to the images. In the first case, only Foreground and
Background seeds are identified from the image. From the seed,
the foreground and background strokes are given as input. In the
second case, the soft constraint is given as input, which allows a
user to draw the strokes to indicate the region that the boundary
should pass through. In the third case, the hard constraint is
given as input, which allows a user to specify the pixels that the
boundary must align with.
The fundamental task of segmenting or partitioning an image
into disjoint regions is used for applications ranging from
medical image analysis, quality control, or military surveillance
and tracking. Although the general segmentation problem
involves separating N distinct partitions, a piecewise
assumption of two sets is generally made. That is, the image is
assumed to be comprised of two homogeneous regions, often
referred to as “object and “Background”. The goal of
segmentation is to accurately capture these regions.
The rest of the paper is organized as follows: We discuss about
the related works in Section 2. We define the proposed work in
Section 3. We present the constrained random walks algorithm
for interactive image segmentation in Section 4. We present the
foreground and background seed generation in Section 5. We
present the soft and hard constrained inputs in Section 6. In
Section 7, we present the results of the proposed framework and
in Section 8, we present our conclusion.
2. RELATED WORKS
There have been many published studies of interactive image
segmentation. Active Contour Models where first proposed by
Kass [7] in 1987 as an Interactive Segmentation method for 2D
images. Active Contour Model treats the desired contour as a
time evolving curve and the segmentation process as an
optimization over time of an adequate energy functional. Live
Wire [10] (also known as intelligent scissors) is a Dynamic
Programming method for extracting a desired object's contour.
The image is treated as a 2D directed weighted graph in which
the nodes are image pixels and the arcs are 8-connectivity
neighbour‟s links. Finding the boundary is implemented as
finding the shorted path in the graph, where low cost path
corresponds to passing near edges and other boundary-related
features. For finding optimal path, Dijkstra's algorithms are
used, which proves the path globally optimal.
Graph Cut segmentation was originally introduced by Boykov
and Jolly [1]. It is shown here as an image segmentation
technique but all arguments hold for N-D as well (in fact, it was
originally introduced as an N-D method). The number of
objects is also arbitrary. The user marks "foreground" and
"background" pixels using free-form, wide-brush strokes. The
system uses the marked pixels as foreground and background
seeds and implies "hard constraints" on them, meaning, the seed
pixels must be labels exactly as the user marked, in the final
segmentation.
Lazy Snapping [8] enables the user to start by specifying
foreground and background marks, and edit the result by border
editing. Furthermore, it attacks the performance problem of
extensive region-based calculation by working first in higher
scale. This is well suited with the dual interaction scheme.
GrabCut algorithm by Rother [12] takes the above Graph Cut
segmentation even further. It applies the min-cut algorithm
iteratively, inferring unknown knowledge from one iteration to
another. This allows it to minimize user interaction by just
specifying the background (and not the foreground) pixels. User
interaction is therefore minimized to placing a rectangle or a
lasso around the object (with respectable distance) instead of
marking brush-strokes.
Simple Interactive Object Extraction (SIOX) is an algorithm for
extracting foreground objects from colour images and videos
with very little user interaction. SIOX works well with noise
and videos [4]. It has been implemented as "foreground
selection" tool in the GIMP. Although the algorithm was
originally designed for videos, virtually all implementations use
SIOX primarily for still image segmentation. The random walks
algorithm [5], [6] model an image as a graph, have been
adopted for various image processing tasks. The random walks
algorithm requires the input of foreground and background
seeds. The random walks algorithm is essentially an approach
that minimizes Dirichlet energy with boundary conditions,
where different boundary conditions results in different
harmonic functions.
The framework proposed in this paper is different from the
above in the sense that it considers a foreground extraction with
minimal user interaction. Further, the objective of our
framework is to extract the foreground object more accurately,
even in camouflage (which allows a visible object to remain
unseen) when background and foreground colour distributions
overlap at least in part.
3. PROPOSED WORK
In the existing system, existing interactive image segmentation
algorithms lack more intelligent ways to understand the
intention of user inputs. Sometimes the user intends to provide
cues for regions while at other times the user intends to focus on
the boundaries. When segmenting a difficult image such as a
cluttered or camouflaged scene, the user frequently struggles
with laborious local editing because the user cannot effectively
communicate his intentions with the underlying segmentation
engine. Even with more and more scribbles, the segmentation
result still may not be what the user wants.
In this paper, we propose a constrained random walks algorithm
that facilitates the use of three types of user inputs: 1)
foreground and background seed input, 2) soft constraint input,
and 3) hard constraint input, as well as their combinations. The

__________________________________________________________________________________________
foreground and background seed input allows a user to draw
strokes to specify foreground and background seeds. The soft
constraint input allows a user to draw strokes to indicate the
region that the boundary should pass through. The hard
constraint input allows a user to specify the pixels that the
boundary must align with. In comparison, the three types of
user inputs and any of their combinations can be supported by
our coherent computational framework consisting of the
constrained random walks algorithm and a local editing
algorithm that imposes soft and hard constraints and allows
more precise contour refinements.
4. CONSTRAINED RANDOM WALKS
In this section, the extension to the random walks algorithm is
proposed. The extension is known as constrained random walks
algorithm to facilitate the use of various user inputs in
interactive image segmentation. The human input could be
valuable in steering the segmentation process in order to obtain
more easily and reliably accurate results according to the visual
perception. The overall design of the proposed interactive image
segmentation method is shown in the Fig. 1.
Fig. 1: Overall design of the proposed method
3.1 Graph Construction
A graph is constructed, in which each node represents a pixel
and neighbouring nodes are connected with undirected edges. A
graph is represented by its vertices and edges as
,vg 
,
where  ivv 
is a set of vertices iv
and jv
and  ije
is
a set of edges ije
bounded by vertices iv
and jv
. The weight
for edge ije
is denoted as ijw
and the degree of node iv is
defined as
 j iji wd
, that is the sum of the weights of all
the edges that incident on iv .
3.2 Edge weights
The edge weights play a critical role since each edge weight
ijw
describes the likelihood of a random walker moving to the
neighbouring node. Each weight should be defined based upon
the distance between two neighbouring pixels/nodes. In the
random walks algorithm, the distance between two nodes is
defined as jiij ggd 
, which is the Euclidean distance in
colour. The prior models have also been incorporated into the
random walks algorithm, where two virtual nodes are added in
the graph to represent the foreground and the background. The
foreground and background node is connected to every pixel
node and the edge weight is assigned to the probability that the
pixel fits the foreground/background prior model. In the
proposed method, the prior models are incorporated directly
into the distance function to avoid the distance connection
problem. The edge weight ijw
and the corresponding distance
are defined as
 2
exp ijij dw  
   222
)Pr()Pr(1 jijiij vvggd  
(1)
The second term
 2
)Pr()Pr( ji vv 
is the prior term, )Pr( iv
denotes the normalized probability that node iv fits the
foreground prior model,  1,0 is a trade off factor, and 
is a scaling factor. The values of
2
ji gg 
and
 2
)Pr()Pr( ji vv 
are normalized to [0, 1] individually.
The foreground and the background are modelled by the
Gaussian mixture model (GMM) and their seeds FS and BS
are used to estimate the model parameters. Let
FviPr
 BviPr
denote the probability that node iv
fits the
foreground and the background GMM. The normalized
probability
)Pr( iv is defined as
BvFv
Fv
v
ii
i
i
PrlogPrlog
Prlog
)Pr(



(2)
Get Seed Input
Foreground and
Background Strokes are
drawn
Soft
Constrained
Input
Segmentation of images
with less accuracy
Hard
Constrained
Input
Segmented image
with more
accuracy

__________________________________________________________________________________________
According to equation (1), it is clear that the second term
 2
)Pr()Pr( ji vv 
should dominate when the foreground and
background colours are well separable. Otherwise, the first term
2
ji gg 
should dominate. Let  be the distance between the
foreground and background GMMs. The Monte-Carlo
simulation is used to approximate the KL-divergence between F
and B.  can be defined as,
 


n
i ii
ii
BvFv
BvFv
n 1 PrlogPrlog
PrlogPrlog1

(3)
Where n is the total number of pixels. After determining the
parameter , the distance measure is calculated. The edge
weight can be applied to graph cuts, random walks, normalized
cuts.
5. FOREGROUND AND BACKGROUND SEED
GENERATION
The foreground and background seed input allows a user to
draw strokes to specify foreground and background seeds. It
identifies the foreground and background object clearly. After
foreground and background seeds are identified and located, the
strokes are drawn. The use of only foreground and background
strokes as input allows the user to segment foreground object
accurately. The design of the foreground and background seed
generation is shown in Fig. 2.
Fig. 2: Design of the Foreground and Background seed
generation
First, the image is given as input. Graph and edge weight is
constructed. The user input includes foreground and
background seeds. Let ip be the probability that an edge
weight will arrive at one of the foreground seeds. Clearly, for
the probability ip , if 1ip , the high probability node is
considered as a Foreground seed. If the probability is
0ip ,
the low probability node is considered as a Background. The
user input includes foreground seeds FS and background
seeds BS , where FS  , BS  , and  BF SS . Let
ip be the probability,
1ip for any Fi S and
0ip for any Bi S . A vertex on which the soft constraint
is imposed has the property that the difference its probability is
½, lies within a small prescribed range [  , ]. A vertex on
which the hard constraint is imposed has a probability of ½.
Let SS
and HS denote soft boundary seeds and hard boundary
seeds, respectively. The Constrained random walks are given by
the following equations:
)(,.
1
HBFij
e
ij
i
i SSSpw
d
p
ij
 


(4)












0,,5.0
,5.0
,0
,1
.,.
 Sii
Hii
Bii
Fii
Svp
Svp
Svp
Svp
ts
The equation (4) can be reformulated into
  



ij Sie Sv
ijiij pppw 22
)5.0()(min








Hii
Bii
Fii
Svp
Svp
Svp
ts
,5.0
,0
,1
.,.
(5)
Where  is a trade off factor controlling the importance of the
difference between the probability of each soft constraint vertex
and ½. Differentiating the objective function of equation (5)
with respect to each ip for
)( HBFi SSSv  and
setting it equal to zero. The constrained random walks solve the
linear system of equation given in equation (6)
Get Input
Image
Graph and Edge
weight is constructed
Compute Probability
High probability node
is taken as Foreground
node
Low probability node
is taken as
Background node
Output:
Background and
Foreground seeds

__________________________________________________________________________________________
 
 











)(,
1
,5.0
1
SHBFij jij
i
Sij jij
i
i
SSSSvpw
d
Svpw
d
p



(6)
The above equation (6) can be considered as adding a virtual
neighbour vertex with probability of ½ to each soft constraint
vertex through a virtual edge with weight  .
The logical step in using Random walks algorithm is to generate
a suitable initialization (seed points). In our approach, a list of
potential seed points are generated, that are used to locate the
foreground and background regions. The seed points are chosen
as the maxima points in the distance map that are located in the
foreground and background. Maxima points are located by
dilating the distance map and finding those points that do not
change in value. It is easy to see that dilation by a mask does
not change the value at only the maxima points. The size of the
mask determines the minimum separation distance between any
two seed-points.
6. CONSTRAINED INPUTS
Two other types of user inputs as constraints into the random
walks algorithm, the extension is known as constrained random
walks. The two constrained inputs are soft constrained input and
hard constrained input.
6.1 Soft Constrained Input
The soft constraint input allows a user to draw strokes to
indicate the region that the boundary should pass through. After
the foreground and background seeds are located, the
foreground and background strokes are drawn. By only
inputting the foreground and background strokes, the weak
boundary between the object and the background cannot be
detected precisely. To overcome this problem, a soft stroke is
drawn. With only one additional soft boundary stroke, which is
partially drawn on the boundary provides better results.
6.1.1 Local Contour Deformation for Soft Constraint
Input
A method of soft constraint input is the local refinement step,
where only additional hard and soft boundary constraints are
allowed to be used to indicate the boundaries. This local
refinement still needs to be performed if there is any soft or
hard boundary stroke being input in the global stage. This is
because the constrained random walks algorithm itself cannot
achieve the purpose of both hard and soft boundary constraints.
For example, the neighbouring nodes of soft constraint have
probability values. Let  be the weighting parameter, if the
 value is too large, the pixels masked by the soft boundary
strokes will have probability values very close to ½, for which
the precise boundary is difficult to locate by thresholding.
For the local refinement, first determine the pixels/positions that
the contour must pass and then build the correspondences
between these pixels and the pixels on the initial contour. After
that, the initial contour is deformed with the correspondences
used as positional constraints and the rest of the pixels on the
initial contour as stay-put constraints.
6.1.2 Optimal Path in Soft Boundary Stroke
Within each soft boundary stroke, an optimal path called soft
boundary path needs to be searched first. The Dijkstra‟s shortest
path algorithm is used to find the shortest path. Let pv
and qv
denote two neighbouring pixels. The local cost pqe
of the
directed edge from pv
to qv
is defined as a weighted sum of
three components: Laplacian zero crossings zf , gradient
magnitude gf
, and the gradient directional cost  qpfd , , i.e,
 
len
qpfqfqf
e
dgz
pq
),(3.0)(6.0)(1.0 

(7)
Where the division by len, denotes the length of the path from
pv
to qv
is to avoid the “shortcut” problem. The weights
specified in equation (7) are empirical values. The purpose of
having the Laplacian zero-crossing term in equation (7) is for
edge localization. The gradient magnitude is to distinguish
between a strong edge and a weak edge. The gradient direction
or orientation term adds a smoothness constraint to the
boundary by associating a relatively high cost for sharp
changes.
6.1.3 Finding Correspondences in Soft Constraint
Input
Correspondences between the soft user input and the initial
contour C obtained in the global stage are established before
contour deformation. The correspondence for a soft boundary
constraint matches the points from the initial contour to the soft
boundary path inside the stroke. First, the two end points of the
soft boundary path are matched to the points on the contour
with minimum geometric distance, and then the in-between
points of the soft boundary path and the contour are matched by
bilinear interpolation. The correspondence set is given as:
   221221 .........,,.........., k
s
k
ss
sssvvv 
(8)
Where is
is the soft boundary constraint point, and
s
iv
is the
corresponding point on the initial contour.

__________________________________________________________________________________________
6.1.4 Refinement by Contour Deformation
The contour deformation can be summarized as: for the initial
contour C, the positional constraints resulted from the
correspondences for the hard and soft boundary points. The
contour deformation is given as:
  j
h
jjj
Cjj
j
s
j
v
hvtsvvjdistfsvLvLv
stay








  
''
'
.,.,)(minarg
2'
22'

(9)
Where  is a trade off parameter for soft boundary constraint
points, and v and
'
v denote two column vectors whose
elements jv
and
'
jv
are the vertices on the initial contour C
and the deformed contour
'
C respectively. The first term
2'
LvLv 
is adapted from the Laplacian mesh processing to
preserve the global shape with the transform matrix L defined
as
ADIL 1
 (10)
Where D is a matrix with iiD =2 and A is the adjacent matrix
defined in equation (11)


 

otherwise
ji
Aij
,0
1,1
(11)
The distance function  )(idistf is considered as the weight.
)(idist is the normalized geometric distance from pixel )(iv
to the input strokes and the function f can be any monotonic
increasing function such that pixels near the input strokes have
small weights and those far away from the input strokes have
larger weights.
6.2 Hard Constrained Input
The hard constraint input allows a user to specify the pixels that
the boundary must align with. In the soft constraint input, the
user draws the strokes along the boundary and the object is
fairly well segmented except some small inaccuracies i.e., the
boundary is not aligned with in the region. To overcome this
problem, the hard boundary constraint input method is
proposed. Boundary pixel selector, which selects pixels on the
desired contour, is introduced as the hard constraint. A vertex
on which the hard constraint is imposed has a probability of ½.
Specifically, by marking five hard boundary pixels, the local
editing algorithm snaps the initial contour to the specified
boundary pixels through local contour deformation, which
results in a smoother and more accurate object contour.
6.2.1 Local Contour Deformation for Hard Constraint
Input
Local Contour Deformation is also performed in hard constraint
input. Deforming contours occur either due to changing region
of partial occlusions or when the object of interest is actually
deforming its shape over a time or space sequence of images.
Higher speed is achieved by the local segmentation determined
only on the vicinity of the route initialization contour and the
segmentation is not performed on the entire input image. The
size of the vicinity of the initialization contour which is taken
into account by the processing of the image is determined
interactively by the user. Terminals of the object and
background in this procedure are determined automatically from
initialization contours specified by the user. The neighbouring
nodes of a hard constraint node have probabilities larger than ½
or less than ½, for which the resulting contour will not pass
through the hard constraint node.
6.2.2 Finding Correspondences in Hard Constraint
Input
Correspondences between the soft user input and the initial
contour C is obtained in the global stage are established before
contour deformation. The correspondence for a hard constraint
point is defined as the point on the contour with minimum
geometric distance. The point from the initial contour to the
hard boundary path inside the stroke is matched. First, the two
end points of the hard boundary path are matched to the points
on the contour with minimum geometric distance. The
correspondence set is given as:
   121121 ..,.........,.........., k
h
k
hh
hhhvvv 
(10)
Where ih
, is the hard boundary constraint point, and
h
iv is the
corresponding point on the initial contour.
7. RESULTS AND DISCUSSIONS
The proposed method has been implemented and tested on
various still images. The proposed system is to accurately
segment the foreground image from the background with the
minimal user interaction using interactive image segmentation.
Figure 3 (a) shows the input image. This image is given as input
for the foreground and background seed generation. The object
with high intensity is considered as the foreground object. The
green seed indicates the foreground object. The blue seed
indicates the background object. Figure 3 (b) shows the
foreground and background seed input in the image.

__________________________________________________________________________________________
(a) (b)
Fig .3(a) Original input image (b) Image with Foreground and
Background Seeds. Green seed indicates the foreground object.
Blue seed indicates the background object.
The foreground and background seeds are generated based on
graph and edge weight calculation. Figure 4 (a) shows the
image with seed input. Figure 4(b) shows the image with
foreground and background strokes. Figure 4 (c) shows the
probability map for the given image. Figure 4(d) shows the
segmented output for the given image.
(a) (b)
(c) (d)
Fig. 4 (a) Image with Foreground and Background Seeds. Green
seed indicates the foreground object. Blue seed indicates the
background object (b) Foreground and Background seeds (c)
Probability map for the Constrained Random walks Algorithm
(d) segmented result for the Foreground and background
strokes.
If the object is in complex background, foreground and
background strokes are drawn on the image. The probability
map provides the good separation between the foreground and
background object. Figure 5 (a) shows the input image. Figure 5
(b) shows the image with soft constraint input. Figure 5 (c)
shows the segmented output for the soft constraint input.
(a) (b)
(c)
Fig. 5 (a) Original input image (b) Soft Constraint input. Red
stroke indicates the foreground object. Yellow stroke indicates
the background object. Green stroke indicates the Soft
constraint input. (c) Segmented result for the soft constraint
input.
The soft constraint input allows the user to draw the strokes to
indicate the region that the boundary must align with. The Soft
Constraint input is given with the boundary brush that roughly
mark parts of the boundary.
Figure 6 (a) shows the hard constraint input. Figure 6 (b) shows
the segmented output.
(a) (b)
Fig. 6(a) hard constraint input, Black pixels are the hard
boundary pixel selector. (b) Segmented Output.
The hard constraint input allows a user to specify the pixels that
the boundary must align with. In the soft constraint input, the
user draws the strokes along the boundary and the object is
fairly well segmented except some small inaccuracies i.e., the
boundary is not aligned with in the region. To overcome this
problem, the hard boundary constraint input is given.

__________________________________________________________________________________________
7.2 Performance Measure
The error rate is defined as the ratio of number of misclassified
pixels to the number of pixels in unclassified region. The error
rate is calculated using the formula:
regionedunclassifiinpixelsno
pixelsiedmisclassifno
.
.

Where misclassified pixels are the incorrect pixels within the
contour region. The unclassified region specifies the pixels that
are present in the contour region. For most of the images in the
MSRC data set, the proposed method achieves very low error
rates. High error rates occur in images where the input
foreground seeds only cover a small portion of the foreground
and, thus, do not cover all the distinct colors of the foreground,
while pixels with similar colors are masked by the background
seeds. For such cases, the GrabCut algorithm and the random
walks algorithm also perform poorly.
Table 1: Performance comparison with different images
In the existing random walks algorithm, the error rate is high,
while for contour deformation the error rate gets decreased. In
the proposed constrained random walk algorithm, the error rate
gets decreased compared to existing algorithm. Therefore the
proposed algorithm is much better than the existing algorithm
which is shown in the Table 1. It is also inferred from the table
that for complex images the error rate gets decreased compared
to the existing method.
CONCLUSIONS AND FUTURE WORK
Interactive image segmentation consists of two components:
Constrained random walks algorithm and local contour
deformation. The proposed method supports multiple intuitive
types of user inputs and also combines the advantages of
different user interactions. The proposed method can be
applicable for camouflaged images and images with thin
structures, if the user inputs are carefully placed.
The foreground and background strokes are carefully drawn at
each side of the object boundary in order to well segment the
object. With a certain amount of user interaction, some
inaccuracies can occur for thin structures whose colour
overlapped with the background colour. In this case, the soft
constraint and hard constraint input is given by the user. The
soft constraint input allows the user to draw soft strokes on the
boundary with the boundary brush. With the soft constraint
input, small inaccuracies occur in the boundary. To overcome
this problem, hard constraint input is provided. The hard
constraint input allows the user to mark the pixels on the
boundary. Specifically, by marking hard boundary pixels on the
inaccurate regions are marked with the boundary pixel selector.
This results in smoother and more accurate segmented output.
The proposed method uses three types of user inputs, and able
to provide more efficient and accurate results. The proposed
method is faster compared to other techniques. Constrained
random walks algorithm requires minimal user interaction to
extract the region. The error rate gets reduced compared to
existing methods.
The proposed method can be extended in two ways. First, its
runtime can be further improved. The response can be improved
through the graphics processing unit. Second, the proposed
method is extended to video segmentation. Extending the
method to video segmentation is not straight forward. This is
because video data consists of thousands of video frames. So
the user has to provide inputs for each video frame. It is also
possible to create a metric to measure the amount and the
complexity of interaction that is required for an interactive
image segmentation to achieve better results.
REFERENCES
[1] Boykov Y. and Jolly M.P., “Interactive graph cuts for
optimal boundary and region segmentation of objects in N-
D images,” in Proc. IEEE Int. Conf. Computer Vision,
2001, pp. 105–112.
[2] Bai X. and Sapiro G., “A geodesic framework for fast
interactive image and video segmentation and matting,” in
Proc. IEEE Int. Conf. Computer Vision, 2007, pp. 1–8.
[3] Falcao A., Stolfi J., and Lotufo R., “The image foresting
transform: Theory, algorithms, and applications,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 26, no. 1, pp. 19–29,
Jan. 2004.
[4] Friedland G., Jantz K., and Rojas R., “SIOX: Simple
interactive object extraction in still images,” in Proc. IEEE
Int. Symp. Multimedia, Dec. 2005, pp. 253–259.
[5] Grady L., “Multilabel random walker image segmentation
using prior models,” in Proc. IEEE Int. Conf. Computer
Vision Pattern Recognition, 2005, pp. 763–770.
[6] Grady L., “Random walks for image segmentation,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 28, no. 11, pp.
1768–1783, Nov. 2006.
[7] Kass M., Witkin .A, and Terzopoulos .D, “Snakes: Active
contour models,” Int. J. Comput. Vis., pp. 321–331,
1988.
Images Error
rate for
Random
Walks
in (%)
Error rate
for Contour
Deformation
in (%)
Error rate
for
Constrained
Random
Walks in
(%)
Statues.jpg 5.85 4.92 4.82
Building.jpg 5.66 4.84 4.75
Penguin.jpg 5.42 4.8 4.63
Tiger.jpg 5.43 4.63 4.45
Boy.jpg 5.38 4.23 4.36
Crow.jpg 5.08 4.16 4.14
Bear.jpg 4.9 4.02 4.08
Flower.jpg 4.5 3.98 3.8

__________________________________________________________________________________________
[8] Li Y., Sun J., Tang C.K. and Shum H.Y., “Lazy snapping,”
ACM Siggraph, pp. 303–308, 2004.
[9] Miranda P., Falcao A., and Udupa J., “Synergistic arc-
weight estimation for interactive image segmentation using
graphs,” Comput. Vis. Image Understand., vol. 114, no. 1,
pp. 85–99, Jan. 2010.
[10] Mortensen E. N. and Barrett W. A., “Intelligent scissors for
image composition,” Comput. Graph. Int. Techniques, pp.
191–198, 1995.
[11] Mortensen E. N. and Barrett W. A., “Interactive
segmentation with intelligent scissors,” Graph. Models
Image Process., vol. 60, no. 5, pp. 349–384, 1998.
[12] Rother C., Kolmogorov V., and Blake A., “Grabcut:
Interactive foreground extraction using iterated graph cuts,”
ACM Siggraph, pp. 309–314, 2004.
[13] Sinop A. K. and Grady L., “A seeded image segmentation
framework unifying graph cuts and random walks which
yields a new algorithm,” in Proc. IEEE Int. Conf. Computer
Vision, 2007, pp. 1–8.
BIOGRAPHIES:
Swathika. P received her UG - B.Tech
(Information Technology) degree in 2009 from
Kamaraj College of Engineering and
Technology. She then completed her P.G -
M.E degree in Computer Science and
Engineering at Mepco Schlenk Engineering
College, Sivakasi in 2011. She is currently
working as an Assistant Professor in Kamaraj College of
Engineering and Technology, Virudhunagar, India. She is a Life
Member of Indian Society of Technical Education (ISTE). Her
area of interest encompasses Digital Image Processing, image
segmentation and Visual cryptography.
Pradhiba Selvarani. M received her UG -
B.Tech (Information Technology) degree in
2009 from Idhaya Engineering College for
Women, chinnasalem. She then completed her
P.G - M.E degree in Computer and
Communication Engineering at National
Engineering College, Kovilpatti in 2011. She is currently
area of interest encompasses Digital Image Processing and
Visual Cryptography.
Arul Mary.S received her UG - B.E (Computer
Science) degree in 2009 from Sri Ramakrishna
Engineering College, Coimbatore . She then
completed her P.G - M.E degree in Computer
Science and Engineering at Mepco Schlenk
Engineering College, Sivakasi in 2011. She is
currently working as an Assistant Professor in Kamaraj College
of Engineering and Technology, Virudhunagar, India. She is a
Life Member of Indian Society of Technical Education (ISTE).
Her area of interest encompasses Digital image processing,
Biometric authentication and Visual cryptography.
Krishnalakshmi. B received her UG - B.E
(Computer Science) degree in 2009 from
P.S.R.Engineering College, Sivakasi. She then
completed her P.G - M.E degree in Computer
Science and Engineering at Anna University of
Technology, Coimbatore in 2011. She had
worked as an Assistant Professor in RVS
educational group of institutions, Coimbatore. She is currently
area of interest encompasses Network Security, Cryptography.

An interactive image segmentation using multiple user inputªs

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to An interactive image segmentation using multiple user inputªs (20)

More from eSAT Journals (20)

Recently uploaded (20)

An interactive image segmentation using multiple user inputªs