A novel predicate for active region merging in automatic image segmentation

IJRET: International Journal of Research in Engineering and Technology ISSN: 2319-1163
__________________________________________________________________________________________
Volume: 02 Issue: 04 | Apr-2013, Available @ http://www.ijret.org 542
A NOVEL PREDICATE FOR ACTIVE REGION MERGING IN
AUTOMATIC IMAGE SEGMENTATION
S.Muthamizhselvi1
, D.Jeyakumari2
, R.Kannan3
1
PG Scholar, 2
Associate Professor, 3
Assistant Professor,
1, 2, 3
Department of ECE, RVS college of Engineering and Technology, Coimbatore, Tamilnadu, India
muthamizhselvi24@gmail.com, dgjeyakumari@gmail.com, kannanvlsi@gmail.com
Abstract
Image segmentation is an elementary task in computer vision and image processing. This paper deals with the automatic image
segmentation in a region merging method. Two essential problems in a region merging algorithm: order of merging and the stopping
criterion. These two problems are solved by a novel predicate which is described by the sequential probability ratio test and the
minimal cost criterion. In this paper we propose an Active Region merging algorithm which utilizes the information acquired from
perceiving edges in color images in L*a*b* color space. By means of color gradient recognition method, pixels with no edges are
clustered and considered alone to recognize some preliminary portion of the input image. The color information along with a region
growth map consisting of completely grown regions are used to perform an Active region merging method to combine regions with
similar characteristics. Experiments on real natural images are performed to demonstrate the performance of the proposed Active
region merging method.
Index Terms: Adaptive threshold generation, CIE L*a*b* color gradient, region merging, Sequential Probability Ratio
Test (SPRT).
-----------------------------------------------------------------------***-----------------------------------------------------------------------
1. INTRODUCTION:
For some applications such as image recognition (or) image
compression, we cannot process the whole image directly for
the reason as it is inefficient and impractical. Therefore we are
going for image segmentation before recognition (or)
compression. Image segmentation is a basic yet still a
demanding problem in computer vision and image processing.
Image segmentation is a key process in object recognition,
target tracking, content-based image retrieval, and medical
image processing, video, and computer vision applications like
medical image analysis such as locating tumors and other
pathologies, in locating objects in satellite images, in face
recognition, fingerprint recognition etc. The goal of image
segmentation is to partition an image into a certain number of
pieces that have coherent features (color, texture, etc.) and, in
the meanwhile, to group the meaningful pieces together for the
convenience of perceiving. In many practical applications, as a
large number of images are needed to be handled, human
interactions involved in the segmentation process should be as
less as possible. For this several general-purpose algorithms
and techniques have been developed for image segmentation.
There is a large amount of literature on automatic image
segmentation. For Example, the spatiotemporal segmentation
is to identify the objects present in the scene represented in a
video sequence. This technique processes two consecutive
frames at a time. Regions are merged based on their mutual
spatiotemporal similarity. Modified Kolmogorov -Smirnov
test for estimating the temporal similarity. This test efficiently
uses temporal information in both the residual distribution and
the motion parametric representation. There are two
complementary graph-based clustering rules are proposed,
namely, the strong rule and the weak rule. The weak rule is
applied after the strong rule. It cannot combine the tracking
information into the segmentation process. The objects
forming the scene are not tracked through time [8]. A hybrid
multidimensional image segmentation algorithm, which
combines edge and region-based techniques through the
morphological algorithm of watersheds. An edge-preserving
statistical noise reduction approach is used as a preprocessing
stage in order to compute an accurate estimate of the image
gradient. Then, an initial partitioning of the image into
primitive regions is produced by applying the watershed
transform on the image gradient magnitude. This initial
segmentation is the input to a computationally efficient
hierarchical region merging process that produces the final
segmentation. The latter process uses the region adjacency
graph (RAG) and nearest neighbor graph (NNG)
representation of the image regions. The final segmentation
provides, due to the RAG, one-pixel wide, closed, and
accurately localized contours/surfaces. It does not involve in
the segmentation of moving 3D images [13].
Different algorithms, based on a combination of two structures
of graph and of two color image processing methods, in order

__________________________________________________________________________________________
to segment color images. The structures used in this study are
the region adjacency graph and the line graph associated. It
shows how these structures can enhance segmentation
processes such as region growing or watershed transformation.
The principal advantage of these structures is that they give
more weight to adjacency relationships between regions than
usual methods [1]. Many tasks in computer vision involve
assigning a label (such as disparity) to every pixel. A common
constraint is that the labels should vary smoothly almost
everywhere while preserving sharp discontinuities that may
exist, e.g., at object boundaries. These tasks are naturally
stated in terms of energy minimization. In this paper, we
consider a wide class of energies with various smoothness
constraints. Global minimization of these energy functions is
NP-hard even in the simplest discontinuity-preserving case.
Two algorithms based on graph cuts that efficiently find a
local minimum with respect to two types of large moves,
namely expansion moves and swap moves. These moves can
simultaneously change the labels of arbitrarily large sets of
pixels. It can find only a local minimum [2].
A database containing „ground truth‟ segmentations produced
by humans for images of a wide variety of natural scenes.
They define an error measure which quantifies the consistency
between segmentations of differing granularities and find that
different human segmentations of the same image are highly
consistent. Use of this dataset is demonstrated in two
applications: (1) evaluating the performance of segmentation
algorithms and (2) measuring probability distributions
associated with Gestalt grouping factors as well as statistics of
image region properties [6]. The technique of scale
multiplication is analyzed in the framework of canny edge
detection. A scale multiplication function is defined as the
product of the responses of the detection filter at two scales.
Edge maps are constructed as the local maxima by
thresholding the scale multiplication results. The detection and
localization criteria of the scale multiplication are derived. At
a small loss in the detection criterion, the localization criterion
can be much improved by scale multiplication. The product of
the two criteria for scale multiplication is greater than that for
a single scale, which leads to better edge detection
performance. [13]
A current level set image segmentation methods, the number
of regions assumed to be known beforehand. As a result, it
remains constant during the optimization of the objective
functional. This study investigates a region merging prior
related to regions area to allow the number of regions to vary
automatically during curve evolution, thereby optimizing the
objective functional implicitly with respect to the number of
regions. The paper gives a statistical interpretation to the
coefficient of this prior to balance its effect systematically
against the other functional terms. They demonstrate the
validity and efficiency of the method by testing on real images
of intensity, color, and motion [4]. The GSEG algorithm is
primarily based on edge detection, dynamic region growth and
multi-resolution region merging [19].
2. PROPOSED METHOD:
The proposed algorithm consists of three steps. First, an edge
detection algorithm is used to produce an edge-map used in
the creation of adaptive gradient thresholds, which selects the
regions of nearby pixels that display similar gradient and color
values, producing an initial segmentation map. Second, a
region growing algorithm is used for growing the regions.
Finally, an active region merging algorithm is used for
merging the similar regions by using a novel predicate P.
Fig -1: Block diagram of the proposed algorithm
2.1 Edge Detection:
The perceived areas with no edges inside them are the
preliminary clusters or seeds selected to initiate the
segmentation of the image. Assuming that an image is a
function f(x, y), the gradient can be defined as its initial
derivative     yfxff  ; . For a vector field f, the
gradient vector to be,
 
   
   










xx
xx
xD
fDfD
fDfD
mnm
n
....................
...................
1
111
(1)
where Djfk is the first partial derivative of the kth component
of f with respect to the jth component of x. For a three-
channel color image, the gradient can be calculated as: let
a,b,c denotes each color channel and x,y denote the spatial
coordinates for a pixel. It is expressed as:
Input image
Color space conversion
Adaptive Threshold Generation
Region Growing
Active Region Merged
image

__________________________________________________________________________________________
222



















dx
dc
dx
db
dx
da
u
(2)



















dy
dc
dx
dc
dy
db
dx
db
dy
da
dx
da
v
(3)
222



















dy
dc
dy
db
dy
da
w
(4)
2.2 Adaptive Threshold Generation:
The Active Region Merging algorithm is started with a color
space conversion of the input image from RGB to CIE
L*A*B. The magnitude of the gradient G(i,j) of the color
image can be calculated by using L*A*B. Primarily, the aim is
to choose a threshold for the initiation of the seed generation
procedure. Preferably, a threshold value could be choosing to
offer the most edges, where as ignoring the noise in images. A
single threshold that may accurately define the boundary of a
given region can permit the other regions to merge wrongly.
So, we are choosing two initial threshold values for the low
and high gradient content in the image. The high initial
threshold can be used for images which have a large gradient
values over a narrow range and a low initial threshold value
can be used for images which have a large gradient values
over a wide range. It guarantees that all significant low
gradient regions are obtained as initial seeds. Low initial
threshold is denoted as λ and high initial threshold is denoted
as λ+5.
Fig -2: Adaptive Threshold Generation
2.3 Seed Generation:
Preliminary seeds are generated by perceiving all the regions
in the image whose gradient values fall below the initial
thresholds λ, and λ+5. If no regions present under this
threshold, the edge value is increased until regions are
identified. The Preliminary Seed Generation makes use of
individual size necessity to select the preliminary seeds, so as
to prevent numerous seed generation within identical and
associated regions. The initial condition is to implement that
seeds to be superior to 0.5% of the image when looking for
regions with a threshold value lower than λ .The basis of this
rule is that frequently backgrounds have false edges produced
by illumination or other varying factors in this gradient range.
The second condition is to implement that seeds to be superior
to 0.25% of the image in the range λ to λ+5, since it is
necessary to differentiate the regions. For distinguishing
purposes the pixels creating each seed obtain an exclusive
label, the mutual set of labeled seeds is called as the Parent
Seeds (PS).
2.4 Region Growing:
The Parent Seeds (PS) grows by increasing the threshold a
single component at a time, for example from 20 to 21. After
each growth, recognition of new regions or child seeds that
falls below the new threshold happens. These child seeds
requires to be categorized in to adjacent-to-existent or non
adjacent seeds. The non adjacent seeds are surplus, because
they can only be added at the starting of an each adaptive
threshold stage. In order to make the region growth method
capable, it is necessary to recognize the parent seed to which
the child is adjacent. The goal is to process all adjacent child
seeds in a vectorized method. To attain this task, we go on to
perceive the outside edges of the map using a nonlinear spatial
filter. The filter is defined as follows:







  
Otherwise
jiPSif
jiPSif
jiF ji
,1
0),(,0
0),(,0
),( , 
(5)
where β is the neighborhood. The result of pertaining this filter
creates the boundary of the PS map. Assign the label for each
pixel in the image. Finally, it produces the region growth map.
2.5 Region Merging:
Region merging is the process of merging the two similar
regions. Similarity is based on color, texture, grey level etc.
Region merging operations eliminates false boundaries and
spurious regions by merging adjacent regions that belongs to
the same object. It controls the over segmentation problem.
Here, the Active Region Merging process occurs by using a
region merging predicate P and cue consistency.
2.5.1 Region Merging Predicate:
The proposed predicate is based on measuring the similarity
between the pixels along the boundary of two regions. The
similarity between the two regions R1 and R2 can be
calculated by using an equation given below:

__________________________________________________________________________________________
    nnRR jiwS ,min, 21 
(6)
where ni and nj are the set of nodes corresponding to the
regions, ni belongs to R1 and nj belongs to R2 and (ni,nj)
belongs to edge E. An edge has a corresponding weight w. We
obtain the similarity between the two neighboring regions by
using a minimal weight edge connecting them. The merging
predicate P will decide whether there is an indication of
merging between the two regions. It requires two features: a
similarity measure, which is used to find out the candidate
region for merging and the consistency property, which
verifies if the regions are homogenous. Merge R1 and R2 if
and only if they are the most similar neighbors.
 
       
 









otherwisefalse
consistentareandb
andSSSaiftrue
P
RR
RRRRRR
RR
ba
.
;,min,min
21
212,1
2,1
(7)
where Ra and Rb is the neighboring region of R1 and R2.
Condition (a) indicates that the connecting edge between R1
and R2 to be the minimal one in either of the neighborhood.
Condition (b) acts as a stopping criterion.
2.5.2 Cue Consistency:
In order to attain the similar regions in region merging, the
proposed predicate P verifies the consistency of the regions.
The region information is obtained by the cues. The selection
of cues can be intensity, color, texture and so on. The
spreading of cues depends on the consistency of the pair of
neighboring regions. In this paper, we are going to use a
sequential probability ratio test method. Assuming that
parameter ϕ is connected to the dissemination of random cues
y. There are two premises involved here: a couple of regions is
“consistent” and “inconsistent”. The consistent premises is
denoted as H0 : ϕ=ϕ0 and inconsistent premises is denoted as
H1 : ϕ=ϕ1. The application of the SPRT to the consistency test
of cues is expressed as follows. The sequence of following
likelihood ratio δj is
 
  nj
yP
yP
j
j
j ........,2,1,log
11
00




(8)
where P0(yj/ϕ0) and P1(yj/ϕ1) are the disseminations of cues.
We are going to use the Gaussian distribution model to fairly
accurate the cue distributions. The two conditional
probabilities are expressed as:
      
      KKXK aKbP
KKXK baKbP
abK
T
babK
T
y
y







1
211
1
100
exp
exp
1 

(9)
where Ka and Kb are the standard color of sampled data in
regions a and b, and Ka+b is the standard value of the samples
union. Xk is the covariance matrix of the regions and λ1 and
λ2 are the scalar parameters. The arrangement of likelihood
ratio is the addition of individual δj, i.e.,
  N
j j1
(10)
where N is the first integer. By using SPRT method, we can
execute the test with a realistic estimation. It can be expressed
as C= log (1-β)/α and D= log β(1-α), where α and β are the
probabilities of decision error given by,
α= Pt {Rejecting H1 when H1 is true}
β= Pt {Accepting H1 when H0 is true}
The selection of α and β gives the region merging feature. α
and β are set to a permanent value of 0.05.
2.5.3 Active Region Merging Algorithm (ARM):
The proposed ARM method deals with the region merging
predicate P. By taking the region merging as an important
problem, the aim is to assign each region a label, such that the
region that belongs to the same object will have the same
label. There are two types of labels for regions: initial label
and final label. The test of consistency and inconsistency is
based on the error probabilities of cue decisions α and β. The
error probabilities are used for finding the homogenous
regions. If any non homogenous regions found, the more
number of tests are used for the decisions. The regions are
arranged in a sequential order from initial label to final label.
It depends on the minimum weight between the two regions.
The whole image is the union of all the regions, F is the sum
of transition costs of overall regions.

R
F
j
jF
(11)
where Fj is the addition of transition costs of Rj.
(a) (b)

__________________________________________________________________________________________
(c) (d)
(e)
Fig-3 (a) Input image (b) Color space conversion (c) Gradient
image (d) Region Growing (e) Region Merging.
3. RESULTS AND DISCUSSIONS:
3.1 Choices for Parameters:
In the proposed ARM method, there are five parameters. They
are q, α, β, λ1, λ2 that controls the consistency premises
evaluation.
Here, q is used to select the amount of data for the random
test. α and β are used for the probability of accepting the
consistency and inconsistency and for rejecting the
consistency and inconsistency. In our method, α and β are set
to a constant value of 0.05 and λ1 ranges from 0.01 to 0.05
and λ2 =1.
3.2 Performance Evaluation:
In our method, we are calculating the accuracy for region
growing and Active region merging. The Accuracy can be
expressed as
%100
N
TNTP
(%)Accuracy 


The terms used to measure the test performance are true
positive (TP), true negative (TN) and total number of images
(N). Results obtained for each test images are given in Table I.
3.3 Comparison of proposed method with GSEG
algorithm:
Here we are going to compare the segmentation results of
ARM with GSEG algorithm. GSEG performs the
segmentation process by using texture model, color edge
detection and cluster formations. There are two parameters
used here. They are texture channel bandwidth and spatial
bandwidth. The segmented image is based on the texture. The
proposed method contains five parameters q, α, β, λ1, λ2.
Compared to GSEG algorithm, the proposed method produces
good localization of boundaries. It reduces the misclassified
regions and other irregularities of the gradient.
Table -1: Region Merging Accuracy by GSEG and ARM for
various images
(a)
(b) (c)
Fig-4 (a) Input image (b) GSEG image (c) ARM image
CONCLUSIONS
In this paper, we proposed a novel predicate for merging the
regions in automatic image segmentation. The Predicate P is
used to determine the evidence of merging between the two
neighboring regions. The Predicate is defined by using
Sequential Probability Ratio Test (SPRT). It produces the high
efficiency and accuracy. It also reduces the misclassified
edges and boundaries.
Image name Region
Growing
Accuracy
(%)
Region
Merging
Accuracy by
GSEG (%)
Region
Merging
Accuracy by
ARM (%)
Tree 85 87 89
Church 93 95 98
Flowers 82 84 88
Man 80 83 85

__________________________________________________________________________________________
ACKNOWLEDGEMENTS
The authors would like to thank their colleagues for their
valuable reviews and comments.
REFERENCES:
[1] A. Trémeau and P. Colantoni, (2000) “Regions
adjacency graph applied to color image Segmentation,”
IEEE Trans. Image Process., vol. 9, no. 4, pp. 735–744,
Apr.
[2] A. Delong, (2010) A. Osokin, H. Isack, and Y. Boykov,
“Fast approximate energy minimization with label
costs,” in Proc. IEEE Conf. Comput. Vis. Pattern
Recognit., San Francisco, CA, pp. 2173–2180 Jun.
[3] B. Paul, L. Zhang, and X.Wu, (2005) “Canny edge
detection enhancement by scale multiplication,”IEEE.
Trans. Pattern Anal. Mach. Intell., vol. 27, no. 9, pp.
1485–1490, Sept.
[4] B. Ayed and A. Mitiche, (2008) “A region merging prior
for variational level set image segmentation,” IEEE
Trans. Image Process., vol. 17, no. 12, pp. 2301–2311,
Dec.
[5] B. Marcotegui and S. Beucher, (2005) “Fast
implementation of waterfall based on graphs,” in Volume
30 of Computational Imaging and Vision. Dordrecht,
Germany: Springer-Verlag, pp. 177–186.
[6] D. Martin, C. Fowlkes, D. Tal, and J. Malik, (2001) “A
database of human segmented natural Images and its
application to evaluating segmentation algorithms and
measuring ecological Statistics,” in Proc. ICCV, pp.
416–423.
[7] D. Comanicu and P. Meer, (2002) “Mean shift: A robust
approach toward feature space analysis,”IEEE Trans.
Pattern Anal. Mach. Intell., vol. 24, no. 5, pp. 603–619,
May.
[8] F. Moscheni, S. Bhattacharjee, and M.Kunt, (1998)
“Spatio-temporal segmentation based on region
merging,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
20, no. 9, pp. 897–915, Sep.
[9] F. Calderero and F. Marques, (2010) “Region merging
techniques using information theory Statistical
measures,” IEEE Trans. Image Process., vol. 19, no. 6,
pp. 1567–1586, Jun.
[10] F. Lecumberry, A. Pardo, and G. Sapiro, (2010)
“Simultaneous object classification and segmentation
with high-order multiple shape models,” IEEE Trans.
Image Process., vol. 19, no. 3, pp. 625–635, Mar.
[11] H. Cheng and Y. Sun, (2000) “A hierarchical approach
to color image segmentation using Homogeneity,” IEEE
Dec.
[12] J. Shi and J. Malik, (2000) “Normalized cuts and image
segmentation,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 22, no. 8, pp. 888–905, Aug.
[13] K. Haris, S. N. Estradiadis, N. Maglaveras, and A. K.
Katsaggelos, (1998) “Hybrid image Segmentation using
watersheds and fast region merging,” IEEE Trans. Image
Process., vol. 7, no. 12, pp. 1684–1699, Dec.
[14] R. Nock and F. Nielsen, (1998) “Statistic region
merging,” IEEE Trans. Pattern Anal. Mach.Intell. vol.
26, no. 11, pp. 1452–1458, Nov.
[15] S.Wang and J.M. Siskind, (2003) “Image segmentation
with ratio cut,” IEEE Trans. Pattern Anal Mach. Intell.,
vol. 25, no. 6, pp. 675–690, Jun.
[16] S. Lee and M. M. Crawford, (2005) “Unsupervised
multistage image classification using Hierarchical
clustering with a Bayesian similarity measure,” IEEE
Mar.
[17] T.Wan, N.Canagarajah, and A.Achim, (2007) “Statistical
multiscale image segmentation via alpha-stable
modeling,” in Proc. IEEE ICIP, vol. 4, pp. IV-357–IV-
360.
[18] Y. Boykov and V.Kolmogorov, (2004) “An
experimental comparison of min-cut/max-flow
algorithms for energy minimization in vision,” IEEE
Trans. Pattern Anal. Mach. Intell., vol.26, no. 9, pp.
1124–1137, Sep.
[19] Luis Garcia Ugarriza, Eli Saber, Sreenath Rao Vantaram,
Vincent Amuso, Mark Shaw, and
RanjitBhaskar,(2009)“Automatic image segmentation by
dynamic region growth and multiresolution merging”,
IEEE Transactions on image processing, vol.18. no.10,
pp.2275 –2288, Oct.
BIOGRAPHIES:
S.Muthamizhselvi received the B.E.
degree in Electronics and communication
engineering from Sree Sowdambika
college of Engineering under Anna
University, Tirunelveli, Tamil Nadu,
India in 2011 and doing M.E. degree in
Communication Systems from RVS
College of Engineering and Technology,
Coimbatore, Tamil Nadu, India. She has presented papers in
the national and International conferences.
D.Jeyakumari received the B.E. degree
in Electronics and communication
engineering from Madurai Kamaraj
University, Madurai, Tamil Nadu, India
in 1997 and M.E. degree in Applied
Electronics from Government College of
Technology, Coimbatore, Tamil Nadu,
and India in 2005. She is doing her Ph.D.
Programme in Anna University, Coimbatore. She has 16 years
of teaching experience. She has published 5 papers in the
reputed international journals and 20 papers in the national
and international conferences. Presently working as an
Associate Professor in ECE Department at RVS College
Engineering and Technology, Coimbatore. She is a life
member of ISTE.

__________________________________________________________________________________________
R.Kannan received the B.E. degree in
Electronics and communication
engineering from Dr.Sivanti Adithanar
college of Engineering, Tiruchendur,
Tamil Nadu, India in 2004 and M.E.
degree in VLSI Design from Sona
College of Technology, Salem, Tamil
Nadu, and India in 2007. He has 7
years of teaching experience. He has
presented papers in the national and international conferences.
Presently working as an Assistant Professor in ECE
Department at RVS College Engineering and Technology,
Coimbatore.

A novel predicate for active region merging in automatic image segmentation

More Related Content

What's hot (17)

Viewers also liked (19)

Similar to A novel predicate for active region merging in automatic image segmentation (20)

More from eSAT Journals (20)

Recently uploaded (20)

A novel predicate for active region merging in automatic image segmentation