A Fast Hybrid Level Set Model For Image Segmentation Using Lattice Boltzmann Method and Sparse Field Constraint

International Journal of Pattern Recognition
and Arti¯cial Intelligence

Vol. 32, No. 6 (2018) 1854015 (24 pages)
#.c World Scienti¯c Publishing Company
DOI: 10.1142/S0218001418540150
A Fast Hybrid Level Set Model for Image Segmentation

Using Lattice Boltzmann Method and Sparse
Field Constraint
Int. J. Patt. Recogn. Artif. Intell. Downloaded from www.worldscientific.com
Dengwei Wang
by RUTGERS UNIVERSITY on 01/12/18. For personal use only.
School of Aeronautics and Astronautics

University of Electronic Science and Technology of China
Chengdu 611731, P. R. China
wdengwei@126.com
Received 25 April 2017

Accepted 31 October 2017
Published 28 December 2017
A novel hybrid-¯tting energy-based active contours model in the level set framework is pro-
posed. The method fuses the local image ¯tting term and the global image ¯tting term to drive
the contour evolution, and a special extra term that penalizes the deviation of the level set
function from a signed distance function is also included in our method, so the complex and
costly reinitialization procedure is completely eliminated. Our model can e±ciently segment
the images with intensity inhomogeneity no matter where the initial curve is located in the
image. In its numerical implementation, two e±cient numerical schemes are used to ensure the
su±cient e±ciency of the evolution process, one is the Lattice Boltzmann Model (LBM), which
is used for breaking the restrictions on time step, the other is the Sparse Field Method (SFM),
which is introduced for fast local computation. Compared with the traditional schemes, these
two strategies can further shorten the time consumption of the evolution process, this allows
the level set to quickly reach the true target location. The extensive and promising experi-
mental results on numerous synthetic and real images have shown that our method can
e±ciently improve the image segmentation performance, in terms of accuracy, e±ciency, and
robustness.
Keywords : Intensity inhomogeneity; level set method; segmentation; lattice Boltzmann model;
sparse ¯eld method.
1. Introduction
Image segmentation is a fundamental task in many image processing and computer
vision applications. Due to the presence of noise, low contrast, and intensity inho-
mogeneity, it is still a di±cult problem in majority of applications.
In the past decades, image segmentation techniques have been extensively and
deeply studied. A well-established class of methods are active contour models,3,11,33
1854015-1
D. Wang
which are based on the theory of surface evolution and geometric °ows, have been
extensively studied and successfully used in image processing. The level set method
(LSM) proposed by Osher and Sethian19 is widely used in solving the problems of
surface evolution. Later, geometric °ows were uni¯ed into the classic energy mini-
mization formulations for image segmentation.14,24–26 Generally speaking, the
existing active contour models can be categorized into two types: edge-based
models2,13,29,32 and region-based models.4,10,16,17,22,31 Each of them has its own pros
and cons.
The edge-based models, usually utilize the gradient information of the image to
construct an edge detector to push the contours toward the desired boundaries of the
objects and ¯nally ensure that the evolution curve converges to these boundaries.
In order to expand the capture scope of the driving force, a balloon force term is often
included in the evolution function, which is used to control the contour to shrink or
expand. In the practice of image segmentation with strong edges, they have achieved
a lot of successful applications. However, they may su®er from some terrible pro-
blems such as level set initialization, edge leakage, and falling into local minimum
value, for image segmentation applications, the existence of these problems will lead
to the segmentation model becoming di±cult to output desired analytical results.
Region-based active contour models use the regional statistical information from
inside and outside of the evolution contour to construct the constraining force to
guide the whole level set evolution process. Compared with the edge-based models,
region-based models have the following obvious advantages: First, region-based
models have more freedom in terms of the contour initialization, i.e. the initial
contour can be located anywhere in the image coordinate system, and the exterior
and interior contours can be detected simultaneously. Second, they are very insen-
sitive to noise and can e±ciently segment the images with weak edges or even
without edges. One of the most successful region-based models is the C–V model,
which has been widely used in binary phase segmentation with the assumption that
each image region is statistically homogeneous. However, the C–V model fails to
segment the images with intensity inhomogeneity.
In order to overcome the segmentation di±culty caused by the intensity inho-
mogeneity, Wang et al.27 proposed a local Gaussian distribution ¯tting (LGDF)
model. The LGDF model utilizes the local region information and thus can provide
accurate segmentation results. However, the ¯nal convergence result of the LGDF
model is related to the initial position of the curve, this means that LGDF model is
sensitive to the initial position of the curve, which greatly limits the scope of its
practical application.
In this paper, we propose a new region-based active contour model which can get
better segmentation results when the intensity of the image is not uniform. In order
to improve the performance of the LGDF model in terms of the degree of freedom of
the initial curve position, we add a global image ¯tting term to the energy functional
of the LGDF model.
1854015-2
After the construction of the active contour model, we need to choose the ap-
propriate numerical solution to solve our model. In order to solve the active contour
model in the level set framework, most classical methods such as the upwind scheme
are based on some ¯nite di®erence, ¯nite volume or ¯nite element approximations
and an explicit computation of the curvature.18 Since only a very small time step can
be taken, thus all of these methods require a large amount of CPU time to complete
their evolution process.
Recently, the lattice Boltzmann method (LBM) has been introduced as an al-
ternative scheme for solving level set equation.30 It can better handle the high
computation problem because the curvature is implicitly computed and the com-
plexity of the algorithm is very low and has a characteristic of high degree of par-
allelism, in addition, due to the use of a very large time step, the evolution process
will be greatly accelerated. In view of the aforementioned advantages, we use LBM to

solve our level set equation.
Most existing numerical methods update the level set function in the full
image domain, which makes them inapplicable in large scale practical problems.
Since the curve evolution is smoothly performed according to the Euler–Lagrange
equation, the use of pixels which are far away from the current contour does not
a®ect the evolution. Therefore, only pixels close to the current contour need to be
considered.
In addition, in view of the Sparse Field Method (SFM)28 having the ability to
signi¯cantly reduce the number of pixels that need to be updated in a single iteration
process, this paper uses SFM as our other important component of the numerical
solution system (one of which is the aforementioned LBM).
The remainder of this paper is organized as follows. Section 2 is a brief description
of the classical C–V and LGDF models, Sec. 3 presents the formulation and imple-
mentation of the proposed model. Section 4 validates the proposed model by ex-
tensive experiments on synthetic and real images. Lastly, conclusions are drawn in
Sec. 5.
2. Background
2.1. C–V model
In the classical active contour model, the external energy is mainly dependent on the
local edge gradient information to detect the potential objects in an image. However,
for images whose boundaries are either smooth or not necessarily de¯ned by gradient,
it will be di±cult to obtain desired results. In order to solve this problem e®ectively,
Chan and Vese4 proposed a new active contour model which was based on the
simpli¯ed Mumford–Shah model, commonly referred to as C–V model. The model
depends on the global information of homogeneous regions. The energy functional is
1854015-3
D. Wang
de¯ned as follows:
Z
EnergyðC; Min ; Mout Þ ¼ 1 ðIðx; yÞ Min Þ 2 dxdy
insideðCÞ
Z
þ 2 ðIðx; yÞ Mout Þ 2 dxdy þ lengthðCÞ
outsideðCÞ
þ areaðinsideðCÞÞ; ð1Þ
where , , 1 and 2 are positive constants, we generally choose the parameters as

follows: 1 ¼ 2 ¼ 1; ¼ 0. Min and Mout are the intensity averages of Iðx; yÞ inside
C and outside C, respectively.
In order to solve this minimization problem, the LSM is introduced which

represents the curve C by the zero level set of a Lipschitz function ðx; yÞ: ! R,
such that ðx; yÞ > 0 if the point ðx; yÞ is inside C, ðx; yÞ < 0 if ðx; yÞ is outside C,
and ðx; yÞ ¼ 0 if ðx; yÞ is on C. Thus, the energy functional Energy ðC; Min ; Mout Þ
can be reformulated in terms of the level set function ðx; yÞ as follows:
Z
Energy" ðC; Min ; Mout Þ ¼ 1 ðIðx; yÞ Min Þ 2 H" ððx; yÞÞdxdy

Z
þ 2 ðIðx; yÞ Mout Þ 2 ð1 H" ððx; yÞÞÞdxdy
Z
þ " ððx; yÞÞjrðx; yÞjdxdy
Z
þ H" ððx; yÞÞdxdy; ð2Þ

where H" ðzÞ and " ðzÞ are, respectively, the regularized approximation of Heaviside
function HðzÞ and the one-dimensional Dirac measure ðzÞ which are de¯ned as
follows:

1; if z 0 d
HðzÞ ¼ ; ðzÞ ¼ HðzÞ: ð3Þ
0; if z < 0 dz
This minimization problem is solved by deducing the associated Euler–Lagrange

equations and updating the level set function ðx; yÞ by the gradient descent method
(with ð0; x; yÞ ¼ 0 ðx; yÞ de¯ning the initial contour):

@ r
¼ " ðÞ div 1 ðI Min Þ 2 þ 2 ðI Mout Þ 2 ; ð4Þ
@t jrj
where Min and Mout can be updated iteratively by
R
Iðx; yÞH" ððx; yÞÞdxdy
Min ðÞ ¼ R ; ð5Þ

Hð" ðx; yÞÞdxdy
R
Iðx; yÞð1 H" ððx; yÞÞÞdxdy
Mout ðÞ ¼ R ; ð6Þ

ð1 Hð" ðx; yÞÞÞdxdy
1854015-4
and 1 ðI Min Þ 2 þ 2 ðI Mout Þ 2 is the global image ¯tting force, which uses the
global image information of the input image to guide the evolution of level set model.
As a representative region information-based level set segmentation model, the C–V
model has an important characteristic, that is, its evolution is not sensitive to the
initial position of the curve. However, when the intensity of the image is not ho-
mogeneous, the di®erence between the two means (Min and Mout ) and the real image
data will be very large, this phenomenon will inevitably lead the C–V model to be
di±cult to give an ideal segmentation result.
2.2. LGDF model

Wang et al.27 proposed a region-based active contour model which considers the local
image information and achieves promising results. They de¯ned the following energy
functional:
Z
LGDF
E LGDF
ð; u1 ; u2 ; 1 ; 2 Þ ¼
2 2
E x ð; u1 ðxÞ; u2 ðxÞ; 1 ðxÞ 2 ; 2 ðxÞ 2 Þdx; ð7Þ

where E xLGDF is de¯ned by
E xLGDF ð; u1 ðxÞ; u2 ðxÞ; 1 ðxÞ 2 ; 2 ðxÞ 2 Þ

Z
¼ !ðx yÞ log p1;x ðIðyÞÞM1;" ððyÞÞdy
Z
!ðx yÞ log p2;x ðIðyÞÞM2;" ððyÞÞdy; ð8Þ
where M1;" ððyÞÞ ¼ H" ððyÞÞ, M2;" ððyÞÞ ¼ 1 H" ððyÞÞ.

8
< 1 exp jdj
> 2
if jdj
!ðdÞ ¼ a 2 2 ð9Þ
>
:
0 if jdj > ;

1 ðui ðxÞ IðyÞÞ 2
pi;x ðyÞ ¼ pffiffiffiffiffiffi exp ; ð10Þ
2i ðxÞ 2i ðxÞ 2
where i ðxÞ and i ðxÞ are local intensity means and standard deviations, respec-
tively, the calculation expressions are as follows:
R
!ðy xÞIðyÞMi;" ððyÞÞdy
ui ðxÞ ¼ R ; i ¼ 1; 2
!ðy xÞMi;" ððyÞÞdy
R ð11Þ
!ðy xÞðu ðxÞ IðyÞÞ 2 M ððyÞÞdy
i ðxÞ 2 ¼ R i i;"
; i ¼ 1; 2:
!ðy xÞMi;" ððyÞÞdy
1854015-5
D. Wang
Minimizing the energy functional by using the steepest descent method, we derive
the gradient descent °ow:
@
¼ " ðÞðe1 e2 Þ; ð12Þ
@t
where
Z
ðui ðyÞ IðxÞÞ 2
ei ðxÞ ¼ !ðy xÞ logði ðyÞÞ þ dy; i ¼ 1; 2: ð13Þ
2i ðyÞ 2
In the above equations, the regularized versions of Heaviside function H and Dirac
function are utilized as follows:
8 z
> 1 2
>
< H" ðzÞ ¼ 1 þ arctan
2 "
ð14Þ
>
> 1 "
: " ðzÞ ¼ ; z 2 R:
"2 þ z2
The parameter " a®ects the pro¯le of " ðÞ. A larger " will lead to a broader pro¯le,
which will enlarge the capture range but decrease the accuracy in the ¯nal contour
location.
In actual calculation, the construction process of the local ¯tted image ðe1 e2 Þ is
based on all the pixels in a local rectangle window, this localization property is the
real reason for the LGDF model to be able to segment nonhomogenous images.
However, when the contour is located at a certain location where e1 ¼ e2 , the local
image ¯tting force will be zero, which leads to the evolution process becoming
trapped into certain local minima, thus the segmentation result has a strong corre-
lation with the initial position of the curve.
3. The Proposed Model

3.1. The formulation of the proposed model
After being inspired by the two models of C–V and LGDF, we de¯ne the energy
functional as follows:
E localglobal ð; Min ; Mout ; u1 ; u2 ; 21 ; 22 Þ
Z
¼ ð1 ÞE LGDF
þ 1 ðI Min Þ 2 H" ðÞdxdy
Z

þ 2 ðI Mout Þ 2 ð1 H" ðÞÞdxdy ; ð15Þ

where is a weight control parameter and its value is located within the
interval ½0; 1.
For more accurate computation involving the level set function and its evolution,
we need to regularize the level set function by penalizing its deviation from a signed
1854015-6
distance function,15 characterized by the following energy functional:

Z
1
P ðÞ ¼ ðjrðx; yÞj 1Þ 2 dxdy: ð16Þ
2
As in typical level set methods, we need to regularize the zero-level set by penalizing
its length to derive a smooth contour which is as short as possible during evolution:
Z
LðÞ ¼ jrHððx; yÞÞjdxdy: ð17Þ

In summary, we can express the total energy functional as follows:

Etotal ¼ a E localglobal þ b P ðÞ þ c LðÞ; ð18Þ
where a, b and c are control parameters to balance the contribution of each

energy term.
Keeping Min and Mout ¯xed, and minimizing the entire energy functional with
respect to , we deduce the associated Euler–Lagrange equation for as follows:
@
¼ a " ðÞð1 ðI Min Þ 2 þ 2 ðI Mout Þ 2 Þ a ð1 Þ " ðÞðe1 e2 Þ
@t
r r
þ b div þ c " ðÞdiv ; ð19Þ
jrj jrj
where Min , Mout , e1 and e2 are formulated by Eqs. (5), (6) and (13), respectively.
The new hybrid ¯tting energy is a weighted linear combination of the global
image ¯tting force from the C–V model and the local image ¯tting force from the
LGDF model. The advantages of this weighted combined form of image ¯tting en-
ergy are as follows: The global image ¯tting force component makes the combined
model insensitive to the initial position of the curve, and the local image ¯tting force
component makes the combined model segment the images with intensity inhomo-
geneity. We combine these two forces together with the control parameter so that
the new hybrid model can have the advantages of the C–V model and the LGDF
model. Therefore, the proposed model can e®ectively deal with the intensity inho-
mogeneity, regardless of the initial contour starting at the image.
The parameter plays an important role to balance the contribution of the
aforementioned two forces, which should be determined based on the degree of inho-
mogeneity of the current image. When there is serious inhomogeneity in the image, we
need to select a small parameter , in this case, the driving force of contour evolution is
mainly from the local image ¯tting force. On the contrary, if the inhomogeneity e®ect is
not obvious, we have to select a bigger parameter in which case the evolution process
of the curve is mainly controlled by the global image ¯tting force.
The proposed hybrid level set model is constructed based on the core processing
ideas of the C–V model and LGDF model. If we set the parameter in formula (17)
to 1 and 0, our model will be degraded to the C–V model and LGDF model, re-
spectively. However, our model takes into account both the local and global image
1854015-7
D. Wang
information, therefore, its segmentation performance is better than the C–V model
and LGDF model in general.
3.2. The implementation of the proposed model

3.2.1. The reasons for the high computational complexity of traditional
level set methods
The traditional level set methods usually need to spend more iterative times
(corresponding to a higher time consumption) to segment an image, which is un-
acceptable for image data-based real-time applications or mass image data proces-
sing problems. The following two reasons led to this high computational complexity
phenomenon:
(1) Constrained iterative time step

An explicit scheme is the most popular way for solving Eq. (18), but due to the
Courant–Friedreichs–Lewy (CFL)34 condition which asserts that the numerical
waves should propagate at least as fast as the physical waves, so the curve can only
move a small distance in each iteration, it requires very small time step and if the
curve is not near the edge of interested object, the curve may take a long time to
reach the ¯nal position.
(2) Full image plane update
A direct implementation of level set evolution equation involves the re-estimation of
the level set function at all pixels, i.e. not simply the zero-level set corresponding to
the current front. This front propagation method is computationally expensive, due
to a large amount of useless calculations that are performed for pixels which are out
of interest during the front propagation.
3.2.2. Two e±cient strategies for accelerating the evolution of level set function
(1) Lattice Boltzmann method for breaking the restrictions on time step
The CFL condition limits the time step of the traditional numerical solution of the
level set equation, which leads to the increase of the number of iterations in the
evolution process. Under the ¯nite di®erence framework, the process needs to ap-
proximate the continuous PDE to a discrete form, while LBM5 derives a continuous
PDE which has the same form of the level set equation from a discrete form. Since the
time step is not strongly restricted and highly parallelizable, LBM is a fast numerical
solution of the level set equation.
LBM is proposed as a computational °uid dynamics (CFD) method for °uid
modeling.21 Instead of solving the Navier–Stokes equations, the discrete Boltzmann
equation is solved to simulate the °ow of the Newtonian °uid by collision models
such as Bhatnagar–Gross–Krook (BGK).6,9
In this paper, we use the D2Q9 (2D with eight links with its neighbors and one
link for the cell itself) LBM lattice structure. Figure 1 shows a typical D2Q9 model.
1854015-8
f 6 , e6 f 2 , e2 f5 , e5
f3 , e3 f1 , e1
f 0 , e0
f 7 , e7 f 4 , e4 f8 , e8
x
Fig. 1. Spatial structure of the D2Q9 LBM lattice.
The evolution equation of LBM can be written as
1 eq D
fi ðr þ ei t; t þ tÞ ¼ fi ðr; tÞ þ ½f i ðr; tÞ fi ðr; tÞ þ 2 F ei ; ð20Þ

bc
where ei is the velocity vector of a given link i, fi ðr; tÞ the distribution of the particle
that moves along that link, t the time, t the time step, r the position of the cell, F
the body force, D the grid dimension, b the link at each grid point and c the length of
each link which is set to 1 in this paper. The parameter
represents the relaxation
time determining the kinematic viscosity in Navier–Stokes equations, and f ieq is the
local equilibrium particle distribution which has the following form:
f ieq ð; uÞ ¼ ðAi þ Bi ðei uÞ þ Ci ðei uÞ þ Di ðuÞ 2 Þ; ð21Þ
where the constant coe±cients Ai to Di are determined based on the geometry of the
lattice links, and u are, respectively, the macroscopic °uid density and velocity
calculated from the particle distributions as
X 1X
¼ fi ; u ¼ fe: ð22Þ
i
i i i
For di®usion problems, the equilibrium function can be simpli¯ed as30

f ieq ð; uÞ ¼ Ai ð23Þ
In the case of D2Q9 structure, the concrete form of is f ieq

9 3
f ieq ¼ i 1 þ 3ðei uÞ þ ðei uÞ 2 ðuÞ 2 ; i ¼ 0; 1; . . . ; 8: ð24Þ
2 2
1854015-9
D. Wang
where 0 ¼ 4=9, 1;2;3;4 ¼ 1=9, 5;6;7;8 ¼ 1=36, at the same time, there is a relation-
ship between the relaxation time
and the di®usion coe±cient :
2
¼ ð2
1Þ: ð25Þ
9
By performing the Chapman–Enskog analysis the following di®usion equation can be
recovered from the LBM evolution equation30:
@
¼ r r þ F : ð26Þ
@t
By replacing with the signed distance function in Eq. (26), since level set
function has the signed distance property jrj ¼ 1, we have the following
expression:

@ r
¼ r r þ F ¼ div þ F; ð27Þ
@t jrj
where \div" is the divergence operator, and the body force F represents the link with
the image data in the LBM solver. Table 1 gives the corresponding forms of
and F
in D2Q9 LBM equation for the C–V model, LGDF model and our model described in
subsection 3.1., the proposed level set equation can therefore be solved using the
following lattice Boltzmann equation:
1 eq
fi ðr þ ei t; t þ tÞ ¼ fi ðr; tÞ þ ½f ðr; tÞ fi ðr; tÞ

i
D
þ ða " ðÞð1 ðI Min Þ 2 þ 2 ðI Mout Þ 2 Þ
bc 2
a ð1 Þ " ðÞðe1 e2 Þ þ b Þ: ð28Þ
Table 1. Corresponding forms of

and f in D2Q9 LBM equation for level set methods.
Method Level Set Equation
and F

C–V model @ r 9
¼ " ðÞ div
¼ " ðÞ þ 0:5;
@t jrj 4
1 ðI Min Þ þ 2 ðI Mout Þ 2
2
F ¼ " ðÞ½ 1 ðI Min Þ 2

þ2 ðI Mout Þ 2
LGDF model @ /
¼ " ðÞðe1 e2 Þ
@t
Our model @ 9
¼ a " ðÞð1 ðI Min Þ 2
¼ ðc " ðÞ bÞ þ 0:5;
@t 4
þ 2 ðI Mout Þ 2 Þ F ¼ a " ðÞð1 ðI Min Þ 2
a ð1 Þ " ðÞðe1 e2 Þ þ 2 ðI Mout Þ 2 Þ

r a ð1 Þ " ðÞðe1 e2 Þ þ b
þ b div
jrj

r
þ c " ðÞdiv
jrj
1854015-10
After adopting the LBM ideology, we do not need to explicitly calculate the cur-
vature since it is implicitly handled by the LBM.
(2) Sparse Field Method for fast local computation
In traditional level set methods, it is required to update the level set function value in
the whole image plane, which greatly increases the CPU time consumption of the
operation process. In order to overcome this drawback, a narrow band approach
which was initially proposed in Ref. 7 and extensively analyzed and optimized in
Ref. 1 is proposed, its key idea is to deal only with pixels which are close to the latest
position of the zero level set in both directions (inwards and outwards), and the
SFM28 takes this strategy to the extreme since it computes the updates on a
band of the grid points that is only one point wide. Therefore, the calculated
amount increases with the size of the curve length, rather than the resolution of
the grid.
The SFM uses lists of points that represent the zero-level set as well as
points adjacent to the zero-level set. By using these lists and carefully moving points
to and from the appropriate list, a very e±cient representation of can be
maintained.
These lists are implemented as doubly-linked lists. These lists have the properties
that elements can be added dynamically, and that elements can be removed from the
middle of the list. This type of data-structure is available in most programming
languages (for example, the vector class in Cþþ).
Five lists are used in the SFM to represent ¯ve di®erent levels:
8
>
> L2 ! ½2:5; 1:5Þ
>
>
>
>
>
> L1 ! ½1:5; 0:5Þ
>
<
L0 ! ½0:5; 0:5
>
>
>
>
>
> L1 ! ð0:5; 1:5
>
>
>
:
L2 ! ð1:5; 2:5:
Each list holds the x, y, and z location of pixels in the image. In addition to the lists,
two arrays are used. The ¯rst is the array. It is the same size as the image domain
and should be maintained at full °oating-point precision. The second array is a label
map the same size as . This label map is used to record the status of each point and
provides a way to look up which list a point belongs to. The label map will only
contain the values f3; 2; 1; 0; 1; 2; 3g.
The procedure of SFM can be divided into two stages: initialization and contour
evolution.
Figure 2 shows an instance of the initialization of SFM. In the level set function
evolution stage, the updated status of L0 is determined by the LBM iteration. Then,
by fusing the neighborhood information, the updated status of points around L0 can
1854015-11
D. Wang
Fig. 2. One example of the initialization in SFM.
be deduced and stored in ¯ve new lists, i.e.

8
> S2 ! points moving to L2
>
>
>
> ! points moving to L1
>
< S1
S0 ! points moving to L0
>
>
>
> S1 ! points moving to L1
>
>
:
S2 ! points moving to L2 :
Finally, points in the above lists update their status and one movement of the
contour is ¯nished. This process is repeated until convergence is reached.
3.3. Algorithm of the fast hybrid level set method for image segmentation
The aforementioned two strategies can e®ectively overcome the problems of high
computational complexity in traditional level set methods. Our fast hybrid level set
algorithm (with the aforementioned acceleration strategies as traction) for non-
homogenous image segmentation is performed as follows:
(1) Initialize the level set function as a signed distance function;

(2) Update Min , Mout , e1 and e2 using (5), (6) and (13), respectively;
(3) Compute the input of our LBM such as
and F according to Table 1 and our
level set equation (19);
(4) Resolve the level set equation using LBM with (28);
1854015-12
(5) Accumulate the values of fi ðr; tÞ at each grid point with Eq. (23), which updates
the values of and locates corresponding contour;
(6) Update the narrow band: By using the new values in evolving curve, update
other points in the narrow band which has been described in Sec. 3.2.2;
(7) Return to step (3) until the evolution process reaches the state of convergence.
4. Experimental Results
In this section, we evaluate the e±ciency of our fast hybrid level set algorithm for
nonhomogenous image segmentation. The experiments are implemented by Matlab
R2012a on a computer with 2.3G Intel Core i7 CPU, 8G RAM, and Windows 7
operating system.
4.1. Comparisons with the C–V model

Figure 3 shows the comparisons of segmentation result between our model and the
C–V model on three synthetic images (through them we can better simulate the non-
homogeneous characteristics). For fairness, we set the same initial curves for the two
models. From these images, we can ¯nd that the nonhomogeneity is very obvious.
As we expected (the description details are located in Sec. 2.1.), the ¯nal convergence
(a) (b) (c)
Fig. 3. Comparisons of our model with the C–V model in segmenting three synthetic images. Column (a)
Initial contour and input image. Column (b) Final contour of the C–V model. Column (c) Final contour of
our model.
1854015-13
D. Wang
of C–V model is very unsatisfactory, however, our model yields accurate segmen-
tation results, all of this is due to the fact that our model combines local and global
image energy e®ectively.
4.2. Comparisons with the LGDF model

In order to verify the initial contour position-insensitive characteristic (please refer to
Sec. 2.2. for the details of the description) of the proposed model, i.e. the ¯nal
segmentation result is independent of the starting position of the initial contour, we
use our model to segment an infrared image with intensity inhomogeneity shown in
Fig. 4.
Figure 4(a) shows the initial contours which have di®erent shapes and initial
positions. Figures 4(b) and 4(c) show the ¯nal locations of the evolution curve by
the LGDF model and our model, respectively. From the ¯nal segmentation
results, we found that the LGDF model outputs an accurate result only when the
initial contour is included inside the object, however, our model yields the same
exact segmentation results under all initialization forms. This fully shows that the
(a)
(b)
(c)
Fig. 4. Comparisons of our model with the LGDF model in segmenting an infrared car image under three
di®erent initialization fashions. Row (a) Initial contour and input image. Row (b) Final contour of the
LGDF model. Row (c) Final contour of our model.
1854015-14
(a)
(b)
(c)
Fig. 5. Comparisons of our model with the LGDF model in segmenting an infrared mouse image under
three di®erent initialization fashions. Row (a) Initial contour and input image. Row (b) Final contour of
the LGDF model. Row (c) Final contour of our model.
segmentation results of this model have nothing to do with the characteristics of

the initial curve.
Figure 5 shows another set of our validation tests, its purpose is similar to
Fig. 4. From this set of experiments, we found a phenomenon similar to Fig. 4, i.e.
only when the initial contour is located inside the target, the LGDF model can
output the correct segmentation results, in sharp contrast to this, our model
yields the same correct segmentation results under all initialization styles. This
further proves that our model has strong initial contour position-insensitive
characteristic.
4.3. Comparisons with the traditional nonlevel-set

segmentation methods
In Secs. 4.1 and 4.2, we compare the model of this paper with the other two-level set
segmentation methods, these methods, however, still belong to the level set type.
In order to broaden the breadth of the comparison process, we also incorporate other
nonlevel-set kind of segmentation methods into our comparison framework. From
the traditional classical methods, we chose the following three as our comparison
objects: EM,8 OTSU,20 PCNN.12 The comparisons of the proposed model with the
three methods are shown in Fig. 6, when we initialize the level set segmentation
1854015-15
D. Wang
(a)
(b)
(c)
(d)
(e)
Fig. 6. Comparisons of our model with the traditional nonlevel-set segmentation methods. Row (a) Input
image and the initial curve required by our model. Row (b) Results of the EM algorithm. Row (c) Results
of the OTSU algorithm. Row (d) Results of the PCNN algorithm. Row (e) Results of the proposed
algorithm.
1854015-16
method, we adopt the manual initialization style, of course, if other initialization

methods are adopted, the result will be the same. In order to conform to the output
form of the level set segmentation, the target contours corresponding to the seg-
mentation results are superimposed onto the original image.
From this set of real image segmentation experiments (as shown in Fig. 6), we ¯nd
that the traditional segmentation methods have di±culties in obtaining ideal seg-
mentation results, and they all have di®erent degrees of defects, however, our method
outputs accurate segmentation results.
We give here a macro explanation for the experimental results in this section:
(1) Because only the intensity information can be used, the statistical modeling
ability of EM algorithm will be a®ected, and then its ¯nal segmentation per-
formance will be a®ected;

(2) The segmentation process of the OTSU algorithm is based on the histogram,
therefore, the spatial information of the image is sacri¯ced, when the target area
of the input image is not uniform, the segmentation result is inevitably deviated;
(3) When the PCNN algorithm performs the segmentation task, it searches pixels
with similar gray levels only within the neighborhood of the internal connection
matrix, in the absence of constraints, some small spurious information will enter
into the segmentation results;
(4) On the contrary, the proposed method is established within the framework of
level set, as a result, we can impose various constraints to ¯lter out the infor-
mation that does not belong to the target, this is an important safeguard
mechanism for accurate segmentation. This is the reason why the output of this
paper is the best.
4.4. Application on real images with noise

The purpose of introducing this set of experiments is to test the anti-noise perfor-
mance of this model. By polluting (perform it through the \imnoise" function of
Matlab) the real images, we obtain this set of noisy test images. In this group of
experiments, the noise added is Gaussian type. The control parameters of imnoise
function are: the mean is equal to 0.01 and the variance is equal to 0.02. Figure 7
presents the results for real-world noisy images. Although the noise information in
the image is very serious, our method still bypasses the noise interference and ¯nally
outputs accurate segmentation results, which show that our method has excellent
anti-noise performance.
4.5. Fast evolution characteristics

In this section, we evaluate the acceleration performance of the two strategies (LBM
and SFM) used in this paper. The test images used here are the same as the images
shown in Figs. 3–5. In order to make our comparative behavior fair, we only compare
1854015-17
D. Wang
Fig. 7. Results of our method for real noisy images. The curve evolution process from the initial contour
(in the ¯rst row) to the ¯nal contour (in the fourth row) is shown in every column for the corresponding
image.
the time consumption of iterative process of the algorithms which have similar ac-
curate segmentation results, the methods involved include: method only with the
local and global driving characteristic (LGD), method with the local and global
driving characteristic and LBM strategy (LGDþLBM), method with the local and
global driving characteristic and SFM (LGDþSFM), our fast level set algorithm
which has all of these two strategies (LGDþLBMþSFM). Table 2 shows the com-
parison results of time cost between the above methods, the sizes of the tested images
are also listed. For the images shown in Figs. 4 and 5, we take the mean value of the
segmentation results.
According to the above comparison results, we can deduce the following conclu-
sions: Firstly, the LBM can decrease the number of iterations of the evolution process
and reduce the total CPU time by several times. Secondly, the acceleration rate of
SFM is even more than one hundred times. Finally, the proposed fast hybrid level set
model which fuses all the two strategies can decrease the iterations from hundreds of
times to dozens of times compared with the level set methods without these two
strategies, at the same time, the total CPU time is reduced from dozens of seconds to
tens of milliseconds.
1854015-18
Table 2. Comparisons between di®erent methods in terms of time cost.

Input Image Algorithm Iterations CPU Time (s) Accelerated Rate
The ¯rst row of column (a) of LGD 60 2.871 1.0

Fig. 3 (100 100 pixels). LGDþLBM 33 1.689 1.7
LGDþSFM 60 0.024 119.3
LGDþLBMþSFM 33 0.018 156.2
The second row of column (a) of LGD 700 21.945 1.0
LGDþSFM 700 0.180 121.7
The third row of column (a) of LGD 200 6.273 1.0
LGDþSFM 200 0.051 123.5

The row (a) of Fig. 4 (118 93 LGD 47 3.289 1.0

pixels). LGDþLBM 20 2.056 1.6
LGDþSFM 47 0.028 115.9
The row (a) of Fig. 5 (303 220 LGD 210 40.045 1.0
LGDþSFM 210 0.333 120.1
The ¯rst column of row (a) of LGD 269 48.000 1.0
LGDþSFM 269 0.388 123.6
The second column of row (a) of LGD 213 37.038 1.0
LGDþSFM 213 0.309 119.9
The ¯rst row of the ¯rst column LGD 100 11.581 1.0
of Fig. 7 (319 127 pixels). LGDþLBM 57 7.238 1.6
LGDþSFM 100 0.101 114.3
The ¯rst row of the second LGD 62 4.088 1.0
column of Fig. 7 (322 254
LGDþSFM 62 0.037 111.9
4.6. Quantitative evaluation

In this section, we use Jaccard similarity (JS)23 to quantitatively evaluate the
performance of the aforementioned methods involved in this paper on images
shown in Fig. 3–5 whose intensity inhomogeneity has di®erent distribution form
and strength. The JS index between two regions Stest and Sground is de¯ned as
JSðStest ; Sground Þ ¼ jStest \ Sground j=jStest [ Sground j, which is the ratio between
the intersectional area of Stest and Sground and the union area of Stest and Sground .
Obviously, the closer the JS value is to 1, the more similar Stest is to Sground . In our
description, Stest is the segmented target region, and Sground is the ground-truth.
1854015-19
D. Wang
Fig. 8. Quantitative comparisons of JS parameters between our model and the C–V model.
In order to ensure the reasonableness of the results, we take the same approach as
Sec. 4.3, i.e. we still take the average value of the results as our calculation objects.
Figure 8 shows the JS values of our model and the C–V model, while Fig. 9 is a
comparison chart of JS parameters between our model and the LGDF model.
Fig. 9. Quantitative comparisons of JS parameters between our model and the LGDF model.
1854015-20
Fig. 10. Quantitative comparisons of JS parameters between our model and the traditional nonlevel-set
segmentation methods.
In Fig. 6, we compared the proposed method with the other three traditional
nonlevel-set segmentation methods. In order to fully demonstrate the superiority of
the proposed method, here, we also give the JS indexes of this set of experimental
results, as shown in Fig. 10.
From these three bar charts, we ¯nd that our model has the best performance
(re°ected by the JS coe±cients) compared with the C–V model, the LGDF, and the
traditional non level-set segmentation methods.
5. Conclusions
This paper presents a fast hybrid level set algorithm for image segmentation. Our
model combines the local and global image energies organically and uses the force
formed by these two energies to promote the evolution of the curve. A large number
of experimental results show that our model can accurately segment the image with
nonhomogeneous properties, and the ¯nal segmentation results are independent of
the initial position of the initial curve. By analyzing the reasons for high computa-
tional complexity, two corresponding strategies named LBM and SFM are induced
to resolve them. The advantage of LBM is that its iterative time step is unrestricted,
therefore, it can greatly reduce the time consumption of evolution process. The use of
SFM can resolve the drawback of full image domain update, and it further reduces
the time consumption of the level set function evolution.
The comparisons with the C–V model, LGDF model, and the traditional non
level-set segmentation methods clearly demonstrate the capability of our method in
handling intensity inhomogeneity and initial curve sensitivity. In addition, our model
also achieves ideal segmentation results on real images containing noise.
In order to verify the acceleration performance of the above two numerical
solutions strategies, we constructed a variety of algorithms and gave the
1854015-21
D. Wang
performance parameters for each algorithm. The results show that the proposed fast
hybrid level set algorithm can reduce the time cost from dozens of seconds to tens of
milliseconds.
Apart from the experiments mentioned above, we also made a complete statistical
analysis of the accuracy of the segmentation results, and gave them in the form of JS
index comparison charts.
Because the LBM and SFM are very suitable for parallel implementation, our
future works can be an implementation of the proposed method using a parallel
device such as the GPU in order to fully take advantage of the LBM and SFM.
Acknowledgments
This work is supported by the Project of the Fundamental Research Funds for the
Central Universities under Grant No. ZYGX2015KYQD032. The authors would like
to thank the anonymous reviewers for their valuable comments and advices.
References
1. D. Adalsteinsson and J. Sethian, A fast level set method for propagating interfaces,
J. Comput. Phys. 118(2) (1995) 269–277.
2. V. Caselles, R. Kimmel and G. Sapiro, Geodesic active contours, Int. J. Comput. Vis.
22(1) (1997) 61–79.
3. V. Caselles, R. Kimmel and G. Sapiro, Geodesic active contours, in Proc. IEEE Fifth Int.
Conf. Computer Vision (1995), pp. 694–699.
4. T. F. Chan and L. A. Vese, Active contours without edges, IEEE Trans. Image Process.
10(2) (2001) 266–277.
5. Q. Chang and T. Yang, A lattice Boltzmann method for image denoising, IEEE Trans.
Image Process. 18(12) (2009) 2797–2802.
6. S. Chen and G. D. Doolen, Lattice Boltzmann method for °uid °ows, Annu. Rev. Fluid
Mech. 30(1) (1998) 329–364.
7. D. Chop, Computing minimal surfaces via level set curvature °ow, J. Comput. Phy.
106(1) (1993) 77–91.
8. A. P. Dempster, N. M. Laird and D. B. Rubin, Maximum likelihood from incomplete data
via the EM algorithm, J. R. Stat. Soc. Series B, 39(1) (1977) 1–38.
9. X. He, Q. Zou, L.-S. Luo and M. Dembo, Analytic solutions of simple °ows and analysis
of nonslip boundary conditions for the lattice Boltzmann BGK model, J. Statist. Phys.
87(1–2) (1997) 115–136.
10. Z. Ji, Y. Xia, Q. Sun, G. Cao and Q. Chen, Active contours driven by local likelihood
image ¯tting energy for image segmentation, Inf. Sci. 301 (2015) 285–304.
11. M. Kass, A. Witkin and D. Terzopoulos, Snakes: Active contour models, Int. J. Comput.
Vis. 1(4) (1988) 321–331.
12. G. Kuntimad and H. S. Ranganath, Perfect image segmentation using pulse coupled
neural networks, IEEE Trans. Neural Netw. 10(3) (1999) 591–598.
13. Q. Li, T. Deng and W. Xie, Active contours driven by divergence of gradient vector °ow,
Signal Process. 120 (2016) 185–199.
14. C. Li, C. Kao, J. Gore and Z. Ding, Minimization of region-scalable ¯tting energy for
image segmentation, IEEE Trans. Image Process. 17(10) (2008) 1940–1949.
1854015-22
15. C. Li, C. Xu, C. Gui and M. D. Fox, Level set evolution without re-initialization: A new
variational formulation, in Proc. IEEE Conf. Computer Vision and Pattern Recognition,
(2005), pp. 430–436.
16. L. Liu, D. Cheng, F. Tian, D. Shi and R. Wu, Active contour driven by multi-scale local
binary ¯tting and Kullback-Leibler divergence for image segmentation, Multimedia Tools
Appl. 76(7) (2017) 10149–10168.
17. S. Niu, Q. Chen, L. deSisternes, Z. Ji, Z. Zhou and D. L. Rubin, Robust noise region-based
active contour model via local similarity factor for image segmentation, Pattern Recognit.
61 (2017) 104–119.
18. S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces (Springer-
Verlag, New York 2003).
19. S. Osher and J. A. Sethian, Fronts propagating with curvature dependent speed: Algo-
rithms based on Hamilton–Jacobi formulations, J. Comput. Phys. 79(1) (1988) 12–49.
20. N. Otsu, A threshold selection method from gray-level histograms, IEEE Trans. Syst.
Man Cybern. 9(1) (1979) 62–66.

21. Y. H. Qian, D. D'Humieres and P. Lallemand, Lattice BGK models for Navier-Stokes
equation, Europhys. Lett. 17(6) (1992) 479–484.
22. R. Ronfard, Region-based strategies for active contour models, Int. J. Comput. Vis. 13(2)
(1994) 229–251.
23. H. Song, Robust visual tracking via online informative feature selection, Electron. Lett.
50(25) (2014) 1931–1933.
24. A. Tsai, A. Yezzi and A. S. Willsky, Curve evolution implementation of the Mumford-
Shah functional for image segmentation, denoising, interpolation, and magni¯cation,
IEEE Trans. Image Proc. 10(8) (2001) 1169–1186.
25. A. Vasilevskiy and K. Siddiqi, Flux-maximizing geometric °ows, IEEE Trans. Pattern
Anal. Mach. Intell. 24(12) (2002) 1565–1578.
26. L. A. Vese and T. F. Chan, A multiphase level set framework for image segmentation
using the Mumford-Shah model, Int. J. Comp. Vis. 50(3) (2002) 271–293.
27. L. Wang, L. He, A. Mishra and C. Li, Active contours driven by local Gaussian distri-
bution ¯tting energy, Signal Process. 89(12) (2009) 2435–2447.
28. R. T. Whitaker, A level-set approach to 3D reconstruction from range data, Int. J.
Comput. Vis. 29(3) (1998) 203–231.
29. C. Xu, and J. L. Prince, Snakes, shapes, and gradient vector °ow, IEEE Trans. Image
Proc. 7(3) (1998) 359–369.
30. Y. Zhao, Lattice Boltzmann based PDE solver on the GPU, Visual Comput. 24(5) (2007)
323–333.
31. S. Zhou, J. Wang, S. Zhang, Y. Liang and Y. Gong, Active contour model based on local
and global intensity information for medical image segmentation, Neurocomput.
186 (2016) 107–118.
32. S. Zhu and R. Gao, A novel generalized gradient vector °ow snake model using minimal
surface and component-normalized method for medical image segmentation, Biomed.
Signal Process. Control 26 (2016) 1–10.
33. G. P. Zhu, Sh. Q. Zhang, Q. S. H. Zeng and Ch. H. Wang, Boundary-based
image segmentation using binary level set method, Opt. Eng. 46(5) (2007) 050501-1–
050501-3.
34. A. P. Zijdenbos, B. M. Dawant, R. A. Margolin and A. C. Palmer, Morphometric analysis
of white matter lesions in MR images: Method and validation, IEEE Trans. Med. Imaging
13(4) (1994) 716–724.
1854015-23
D. Wang
Dengwei Wang re-

ceived his B.S. degree in
Automation from Wuhan
University of Science and
Technology, Wuhan,
China, in 2005, his M.S.
degree in Control Theory
and Control Engineering
from Huazhong Universi-
ty of Science & Technol-
ogy, Wuhan, China, in
2007, and his Ph.D. in Control Science and En-
gineering from Huazhong University of Science &
Technology, Wuhan, China, in 2011. From
September 2011 to October 2014, he was with
the Key Laboratory of Electronic Information

and Control, Southwest China Research Insti-
tute of Electronic Equipment, where he was on

the research sta® and then promoted to Senior
Engineer. He is currently a Lecturer with the
Department of Aeronautics and Astronautics,
University of Electronic Science and Technology
of China. Currently, he is the main teacher of the
undergraduate's course called \Digital Image
Processing". His research interests include visual
target detection and recognition, image seg-
mentation, visual scene understanding, probabi-
listic graph model, variational level set method
and design of software system based on Matlab
GUI architecture.
1854015-24

A Fast Hybrid Level Set Model For Image Segmentation Using Lattice Boltzmann Method and Sparse Field Constraint

Uploaded by

Copyright:

Available Formats

You might also like

A Fast Hybrid Level Set Model For Image Segmentation Using Lattice Boltzmann Method and Sparse Field Constraint

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Fast Hybrid Level Set Model For Image Segmentation Using Lattice Boltzmann Method and Sparse Field Constraint

Uploaded by

Copyright:

Available Formats

International Journal of Pattern Recognition

and Arti¯cial Intelligence

A Fast Hybrid Level Set Model for Image Segmentation

School of Aeronautics and Astronautics

Received 25 April 2017

will be greatly accelerated. In view of the aforementioned advantages, we use LBM to

where , , 1 and 2 are positive constants, we generally choose the parameters as

In order to solve this minimization problem, the LSM is introduced which

This minimization problem is solved by deducing the associated Euler–Lagrange

2.2. LGDF model

where E xLGDF is de¯ned by

E xLGDF ð; u1 ðxÞ; u2 ðxÞ; 1 ðxÞ 2 ; 2 ðxÞ 2 Þ

where M1;" ððyÞÞ ¼ H" ððyÞÞ, M2;" ððyÞÞ ¼ 1 H" ððyÞÞ.

3. The Proposed Model

distance function,15 characterized by the following energy functional:

In summary, we can express the total energy functional as follows:

where a, b and c are control parameters to balance the contribution of each

3.2. The implementation of the proposed model

(1) Constrained iterative time step

Fig. 1. Spatial structure of the D2Q9 LBM lattice.

The evolution equation of LBM can be written as

For di®usion problems, the equilibrium function can be simpli¯ed as30

Table 1. Corresponding forms of

F ¼ " ðÞ½ 1 ðI Min Þ 2

Fig. 2. One example of the initialization in SFM.

be deduced and stored in ¯ve new lists, i.e.

(1) Initialize the level set function as a signed distance function;

4.1. Comparisons with the C–V model

(a) (b) (c)

4.2. Comparisons with the LGDF model

segmentation results of this model have nothing to do with the characteristics of

4.3. Comparisons with the traditional nonlevel-set

method, we adopt the manual initialization style, of course, if other initialization

formance will be a®ected;

4.4. Application on real images with noise

4.5. Fast evolution characteristics

Table 2. Comparisons between di®erent methods in terms of time cost.

The ¯rst row of column (a) of LGD 60 2.871 1.0

LGDþSFM 200 0.051 123.5

The row (a) of Fig. 4 (118  93 LGD 47 3.289 1.0

4.6. Quantitative evaluation

Man Cybern. 9(1) (1979) 62–66.

Dengwei Wang re-

the Key Laboratory of Electronic Information

tute of Electronic Equipment, where he was on

You might also like

E xLGDF ð; u1 ðxÞ; u2 ðxÞ; 1 ðxÞ 2 ; 2 ðxÞ 2 Þ

F ¼ " ðÞ½ 1 ðI Min Þ 2

The row (a) of Fig. 4 (118 93 LGD 47 3.289 1.0