Base Paper
Base Paper
Base Paper
--------------------- (3.1)
Where { }
1
N
k
k
x
=
the feature vectors for each pixel are, { }
1
c
i
i
v
=
are the prototypes of the clusters and the array [ ]
ik
u U =
represents a partition matrix, namely
1
1| 0 1, 1, 2, 3....
c
ik ik
i
u u k N
=
= =
----------- (3.2)
and
1
0
N
ik
k
u N
=
----------------- (3.3)
The parameter p is a weighting exponent on each fuzzy
membership and determines the amount of fuzziness of the
resulting classification. The FCM objective function is
minimized when the high membership values are assigned to
pixels whose intensities are close to the centroid of their
particular class, and low membership values are assigned
when the pixel data is far from the centroid.
The constrained optimization could be solved using one
Lagrange multiplier
2
1 1 1
(1 )
c N c
p
m ik k i ik
i k i
F u x v u
= = =
= +
--------- (3.4)
Where denotes a Lagrange multiplier. The derivative of
m
F w.r.t
ik
u was computed and the result was set to zero,
for 1 p >
2
1 p m
ik k i
ik
F
pu x v
u
------------- (3.5)
1
1
2
1
1
m
ik
m
k i
u
m
x v
| |
=
|
\
----------------- (3.6)
The identity constraint
1
1
c
jk
j
u
=
=
=
| |
=
|
\
---------- (3.7)
This allows us to determine the Lagrange multiplier
IEEE INTERNATIONAL CONFERENCE ON ADVANCES IN ENGINEERING, SCIENCE AND MANAGEMENT
IEEE-ICAESM 2012
1
1
2
1
1
1
1
m
c
j
m
k j
m
x v
=
| |
=
|
\
---------- (3.8)
The zero-gradient condition for the membership estimator can
be rewritten as
2
1
1
1
ik
p
c
k i
j
k j
u
x v
x v
=
=
| |
|
|
---------------- (3.9)
As no constraints, the computations of the prototypes were
straightforward, The minimum of J was computed
with respect to
i
v , and yielded the following equation
0
i
v
J = ----------------- (3.10)
The detailed solution depends on the distance function.
In the case of the Euclidean distance, this leads to the
expression:
1
2 ( ) 0
N
p
ik k i
k
u x v
=
=
------- (3.11)
So the following could be immediately obtained
1
1
N
p
ik k
k
i N
p
ik
k
u x
v
u
=
=
=
----------- (3.12)
The FCM algorithm for segmenting the image into different
clusters can be summarized in the following steps:
FCM Algorithm:
Step 1: Select initial class prototype{ }
1
c
i
i
v
=
.
Step 2: Update all memberships
ik
u with Eq. (3.9).
Step 3: Obtain the prototype of clusters in the forms of
weighted average with Eq. (3.12).
Step 4: Repeat step 2-3 till termination. The termination
criterion is
new old
V V .
Where . is the Euclidean norm. V is the vector of cluster
centers is a small number that can be set by user (here
=0.01).
V. THE MODIFICATION TO THE LEVEL SET METHOD
The level set method was invented by Osher and Sethian [3] to
hold the topology changes of curves. A simple representation
is that when a surface intersects with the zero plane to give the
curve when this surface changes, and the curve changes
according with the surface changes. The heart of the level set
method is the implicit representation of the interface. To get
an equation describing varying of the curve or the front with
time, we started with the zero level set function at the front as
follows:
( ) 0 , , = t y x , if ( ) 1 , y x .( 4.1)
Then computed its derivative which is also equal to zero
0 . . =
t
y
y t
x
x t
.( 4.2)
Converting the terms to the dot product form of the gradient
vector and the x and y derivatives vector, we go
0 . . =
|
|
\
|
\
|
t
y
y t
x
x t
..( 4.3)
Multiplying and dividing by and taking the other part to be
F the equation was gotten as follows:
0 = +
F
t
.(4.4)
According to literature [9]11], an energy function was defined:
( ) ( ) ( )
ext
E E E + =
int
.(4.5)
Where ( )
ext
E was called the external energy, and ( )
int
E
was called the internal energy. These energy functions were
represented as:
( ) ( ) dxdy E
2
int
1
2
1
=
. (4.6)
( ) ( ) ( )
g g ext
vA L E + = .. (4.7)
( ) dxdy g L
g
= (4.8)
( )dxdy gH A
g
= . (4.9)
I G
g
* 1
1
+
= . (4.10)
Where ( )
g
L was the length of zero level curve of ; and
g
A
could be viewed as the weighted area; I was the image and g
was the edge indicator function. In conventional(traditional)
level set methods, it is numerically necessary to keep the
evolving level set function close to a signed distance
function[14][15]. Re-initialization, a technique for
occasionally re-initializing the level set function to a signed
distance function during the evolution, has been extensively
used as a numerical remedy for maintaining stable curve
evolution and ensuring desirable results.
From the practical viewpoints, the re-initialization process can
be quite convoluted, expensive, and has subtle side effects
[16]. In order to overcome the problem, Li et al [8] proposed a
new variational level set formulation, which could be easily
implemented by simple finite difference scheme, without the
need of re-initialization. The details of the algorithm are in the
IEEE INTERNATIONAL CONFERENCE ON ADVANCES IN ENGINEERING, SCIENCE AND MANAGEMENT
IEEE-ICAESM 2012
literature [8]. However, because only the gradient information
was imposed in the edge indicator function, Lis method has a
little effect on the presence of fuzzy boundaries.
In the paper, a innovative method was proposed to modify the
algorithm. The original image was partitioned into some sub
images by FCM. The fuzzy boundary of each sub image was
weighted by , the edge indicator function was redefined:
2
'
g g g + = . (4.11)
Where
1
2
* 1
1
I G
g
+
=
1
I Was the image after clustering. The iterative equation
of level set functional was:
( )
( ) ( )
' '
1
vg g div div
n n
+
|
|
\
|
+
(
(
|
|
\
|
(4.12)
Taking
2
'
g g g + = into 4.12
( )
( )
( )
( )
(
(
(
(
+
|
|
\
|
+ +
|
|
\
|
(
(
|
|
\
|
+ =
+
2
2
1
vg
g div
vg
g div div
n n
(4.13)
Where [ ] 1 , 0 . When processing images of weak
boundary or low contrasts, a bigger was taken; otherwise, a
smaller was taken.
VI. THE GENERATION OF INITIAL CONTOURCURVE
On the basis of FCM clustering in brain MR image
segmentation, the over segmentation usually exists. In this
paper, the result of FCM was used as initial contour curve, and
the automated initialization of deformation model was
finished. For all the pixels in each cluster i.e. white matter, if 4
neighborhoods included the heterogeneous pixel, the pixel was
regarded as candidate boundary point. So the algorithm of
curve tracing [17] was proposed. The exterior boundary of the
cluster was tracked in the candidate boundary points. Finally,
the closed curve was obtained. The candidate boundary points,
whose Euclidean distances to the origin coordinates were
shortest, were chosen as initiation points of curve tracing. The
steps of image segmentation with adapted level set method
were as follows:
Step1. Set the number of clusters, then the original image was
processed with FCM, and calculate the
2
g .
Step2. Choose one cluster, evaluate the inside area
with and the outside area with + , is a plus constant.
The boundary of the area is set to 0. The region of interest is
defined initial contour.
Step3. Minimize the overall energy functional with 4.13
formula.
VII. EXPERIMENTAL RESULTS
The segmentation of image takes an important branch in the
surgery navigation and tumor radiotherapy. In the experiment,
The image data from the IBSR(Internet Brain Segmentation
Repository) are included to test the accuracy and efficiency of
the proposed algorithm. The output of watershed
transformation algorithm figure 1 is as shown below. Firstly
the original MRI brain image is transformed to a proposed
watershed algorithm is that a superimposed image of ridge
lines and original binary image, note the over segmentation.
The approximate contour of white matter was got by FCM
algorithm shown in Figure c of figure 1. The snooping of
regions else appear as a result of the in excess of
segmentation. The initial evolution curve was obtained by the
automated initialization.
With the enhanced method, the curve was successfully
evolved to the hollow white matter boundaries, but only to the
approximately white matter boundaries with Lis method. At
the same time, because the curve has been converged to the
narrow region the object boundaries extraction could not be
implemented with Lis method. But the enhanced method
solved this problem better. On the similar computing proposal,
under a 3.0GHz Pentium iv PC with 1 GB RAM on board, the
average processing time of improved method was 9.6s, and
that was 30.3s with Lis method. The evolution time was
greatly reduced.
VIII. DISCUSSIONS
The need of the re-initialization is completely eliminated by
the proposal of Chunming Li, for pure partial differential
equation driven level set methods, the variational level set
methods. It can be easily implemented by using simple finite
difference method and is computationally more efficient than
the traditional level set methods. But, in this algorithm, the
edge indicator has little effect on the low contrast image.
Meanwhile, the initial contour of evolution needs to be
determined by manual, and it has the shortcomings of time-
consuming and user intervention.
In this paper, we projected a new method to transform the
algorithm. The original image was partitioned with FCM, and
the controlled action of the edge indicator function was
increased. The result of FCM segmentation was used to obtain
the initial contour of level set method. With the new edge
indicator function, results of image segmentation showed that
the improved algorithm can exactly extract the corresponding
region of interest.
IX. CONCLUSIONS
In conclusion, the results of this study confirmed that the
combination of FCM with the level set methods could be used
IEEE INTERNATIONAL CONFERENCE ON ADVANCES IN ENGINEERING, SCIENCE AND MANAGEMENT
IEEE-ICAESM 2012
for the segmentation of low contrast images, such as MR
images. The method has the advantages of no reinitialization,
automation, and reducing the number of iterations. The
validity of new algorithm was verified in the process of
exacting tissues of the brain. In the future research, the effect
of priori information on the object boundary extraction with
level set method, such as boundary, shape, and size, would be
further analyzed. At the same time, the performance of image
segmentation algorithms would be improved by reconstruction
of classic velocity of level set method.
In the future research, noise was added in images prior
information on the object boundary extraction with level set
method, such as boundary, shape, and size, would be further
analyzed. At the same time, the performance of image
segmentation algorithms would be improved by modernization
of classic velocity of level set method.
REFERENCES
[1] J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function
Algorithms. Plenum Press, New-York, 1981.
[2] K. L. Wu, and M.S. Yang, Alternative c-means clustering algorthims,
Pattern Recognition VoI.35, pp.2267-2278,2002.
[3] Osher, S., and Sethian J.A., Front Propagating with Curvature-
Dependent Speed: Algorithms Based on Hamilton-Jacobi Formulations,
Journal of Computational Physics, 79, pp. 12-49, 1988.
[4] Malladi,R.,Sethain,J. and Vemuri,B., Shape modelling with front
propagation: A level set approach. IEEE Trans.Pattern Analysis and
Machine .Intelligence, pp.158-174,1995.
[5] L. Staib, X. Zeng, R. Schultz and J. Duncan. Shape Constraints in
Deformable Models. In Handbook of Medical Imaging, I. Bankman
(ed.), Academic Press, chapter 9, pp 147-157, 2000.
[6] Leventon.M, Faugeraus.O, Grimson.W, and Wells.W, Level set based
segmentation with intensity and curvature priors. IEEE Workshop on
Mathematical Methods in Biomedical Image Analysis Proceedings,
pp.4-11,2000.
[7] Paragios.N, Deriche. R, Geodesic active contours and level sets for the
detection and tracking of moving objects. IEEE Transaction on pattern
Analysis and Machine Intelligence, pp.266-280 Mar 2000.
[8] Vese .L .A, Chan .T. F, A multiphase level set frame wor for image
segmentation using the mumford and shah model. International
Journal of Computer Vision, vol 50 no 3,pp.271-293,2002.
[9] Shi, Y, Karl W. C., Real-time tracking using level set, IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition,pp.34.42,2005.
[10] Li. C., Xu. C., Gui. C., and Fox. M .D, Level set evolution without re-
initialization: A New Varitional Formulation, IEEE Computer Society
Conference on Computer Vision and pattern Recognition, pp.430-436,
2005.
[11] Sethain. l, Level set Methods and Fast Marching Methods . Cambridge
University Press, 1999.
[12] Dunn J.C., A fuzzy relative of the ISODATA process and its use in
detecting compact well-separated Clusters.Journal of Cybernetics vol 3
issue 3, pp.32-57,1973.
[13] Bezedek J., A convergence thheorem for the fuzzy ISODATA
clustering algorthims. IEEE Trans.Pattern Analysis and Machine
Intelligence,pp 78-82,1980.
[14] S.Osher and R.Fedkiw, Level set methods and Dynamic implicit
surfaces,Springer,vol 57, issue 3, pp.112-113,2002.
[15] D.Peng,B.Merriman,S.Osher,H.Zhao, and M.Kang, A PDE- based fast
local level set method, Journal of Computational Physics,vol 155,issue
2,pp.410-438,1996.
[16] J.Gomes and O.Faugeras, Reconciling distance functions and Level
Sets, J.Visual Communic. And Image Representation,vol 11, pp.209-
223,2000.
[17] T. McInerney and D. Terzopoulos, "Deformable models in medical
image analysis: a survey," Medical Image Analysis, vol. 1, no. 2, pp. 91-
108, 1996.
[18] Dao-Qiang Zhang, Song-CanChen, Clustering incomplete data using
Kernel-based fuzzy c-means algorithm. Neural Processing Letters
(2003). Vol 18, Issue 3, pp. 155-162, 2003.
[19] F. Meyer, S. beucher, Morphological Segmentation .J of Visual
Commincation and Image Representation. Vol 1, pp. 21-26,1990.
[20] Sijbers, J., Verhoye, .M., Scheunders, P., A.Vander L., Dyck,D.V.,
Raman, E., Watershed-based segmentation of 3D MR data for volume
quantization. MagneticResonance Imaging, Vol 15(6),679-688,1997.
[21] Vincent, L., Soille, P., Watersheds in digital spaces: an efficient
algorithmbased on immersion simulations. IEEE Transactions on
PatternAnalysis and Machine Intelligence,Vol 13(6), pp. 583-589,1991.
[22] Gonzlez R and Woods R ,Digital Image Processing , Second
Edition,Publishing House of Electronics Industry, pp. 505-507, 2003.
Authors Bio-Data:
Tara.Saikumar was born in 1987 in Hyderabad, India. He received
the B.Tech. degree in Electronics and Communication Engineering
from Joginpally .B.R. Engineering College, Jawaharlal Nehru
Technical University, Hyderabad in 2008. and the M.Tech. degree
from Kakatiya Institute of Technology and Science (KITS),
Warangal, India, in 2010 . He is the university topper in his M.Tech,
KITS, Warangal and also received gold medal from Dr.D.N.Reddy
vice-chancellor JNTUH in his B.Tech for his excellence2004-2008.
He has more than 35 papers in his credit, International journal,
International and National Conferences. He is a life member of
IACSIT, and UAEEE. He got Best paper award in 2011.
He is currently a Asst Professor with the Dept of Electronics and
Communication Engineering, CMR Technical Campus, Hyderabad.
His research interests include digital image processing, digital signal
processing, communication and neural network.
M.Nagarani received her B.Tech, degree in Electronics &
Communication Engineering from JNTUH. M.Tech., degree from
JNTUA. Presently working as Assistant Professor in ECE, VITS,
Karminagar, India. Her research interests include digital image
processing, digital signal processing.
K.Yojana received her B.Tech, degree in E&IE from Vagdevi
College of Engineering, Warangal. M.Tech., degree from KITS
Warangal. Presently working as Assistant Professor in ECE, SBITS,
Khammam, India. Her research interests include image processing, &
signal processing.
B.Shashidhar Working as Asst.professor in SMCET, Hyderabad,
Depatment of Mathematics.He obtained B.Sc, M.Sc. B.Ed. and he
has a vast experience in various engineering college. Life member of
ISTE.
IEEE INTERNATIONAL CONFERENCE ON ADVANCES IN ENGINEERING, SCIENCE AND MANAGEMENT
IEEE-ICAESM 2012
Adaptive threshold
test Image
Watershed
image
Watershed
image
Adaptive threshold
test Image
Adaptive threshold
test Image
Watershed image
Adaptive threshold
test Image
Watershed image
Figures: a b c d
Figure a: Adaptive Thgreshold Test Images
Figure b: WatershedImages
Figure c: FCM Segmentd Images
Figure d: Proposed Level Set Method