201 5 International Conference on Intelligent Computing and Internet of Things (IClT)

Watershed Image Segmentation Algorithm Base on

Particle Swarm and Region Growing

SUN Hui-jie
College of Computer Science and Information Engineering
Harbin Normal University
Harbin Heilongjiang 150025,
China
Sunh858@163.com

Abstract-An improved watershed image segmentation

algorithm is proposed to solve the problems of noise-sensitivity II. WATERSHED ALGORTHM AND REGION GROWING
and over-segmentation. The new algorithm which combined
region growing with classical watershed algorithm is A. Watershed Algorithm
established by constructing an objective function, the Watershed algorithm [81 is based on the simulation of the
parameter of region growing is ensured based on Shannon flooding process implementation, that is the foundation of
entropy. The particle swarm optimization algorithm is 2D plus the height of 3D foundation. There are some local
employed to search global optimization of the objective
minimum point in the image. If a hole is played in each tiny
function. Experimental results show that the new watershed
point, water will slowly into the surface from these holes,
image segmentation algorithm can solve effectively the
starting from the lowest minimwn point, then water
problem of over-segmentation and turns out to be an efficient,
accurate and applied image segmentation algorithm. inundated water basin from the lowest point. When water
surface from two different minimwn point gather together, a
Keywords-particle swarm optimization; image segmentation; dam is built here. At the end of the flush process, each
region growing; mathematical morphology; watershed algorithm minimum point is surrounded by a corresponding water
basin of the dam, the set of whole dam forms a watershed.
Different water basin is made up different region of the
I. INTRODUCTION
image, and the image segmentation is realized.
Image segmentation is divided into several regions of
different characteristics, then the interested target is Experiments show that the results of gradient image on
extracted, it is an important part of the image processing. watershed segmentation is more accurate than in the original
Such as threshold segmentation, region growing image image segmentation, so the gradient image is got with the
segmentation, watershed segmentation algorithm, clustering morphological gradient operator for the watershed
algorithm, neural network segmentation algorithm, they are transformation.
some classical algorithms. GRAD(j) =(j(J)B)-(f0B) (1)
Watershed algorithm is a kind of segmentation Formula 1 is a common calculation method of the
algorithm based on mathematical morphology, it is gradient. In 1997, Wang [9] and proposed a multi-scale
composed of Digabel and Lantujoul in[ll, it is mainly used in morphological gradient operator
the analysis of simple binary image. In order to get more
general image segmentation model; Beucher, Vincent [2Ietc. 1 "n
GRAD(j)= - L.. ,=I ((j(J)B,)-(f0B;))0B'_1 (2)
established theory and algorithm for gray image n
segmentation of watershed. At present, there are many Here: GRAD is morphological gradient, Bj is a structure
improved algorithm based on watershed[31, These algorithms element, and Bj can be any form if Bo<:;;;B 1<:;;; ...<:;;;Bn.
can be classified into three categories: 1) the fuzzy set
theory is used in watershed algorithm of image
B. Region Growing
preprocessing[4]and watershed transform[5];2)a watershed
segmentation in mark image[61;3)the wavelet transform with Region growing is a classical image segmentation
watershed al�orithm is improved to eliminate over­ methods, the basic thought is that the region is made up of
segmentation I. In this paper, a new image segmentation qualitative similarity pixel set. Specifically a seed pixels is a
algorithm is put forward for combining with the region beginning growth point for each separate area firstly (seed),
growing principle and based on optimization method of seeds point will be the same seed pixels for a neighborhood
watershed improved algorithm. or qualitative similarity pixels incorporated into the seed
pixel region. The pixel as the new seed point continue above
process, until no pixel for the conditions of satisfaction. So

978-1-4799-7534-1 114/$31.00 ©201 51EEE 51 Harbin, China

 an area was growing. This process can be realized by qua­
tree algorithm quickly. (5)

III. THE WATERSHED SEGMENTA nON ALGORITHM BASED ON

REGION GROWING
(6)
A. The Watershed Algorithm Based on Region Growing
In this paper, the region growing algorithm is combined
with the watershed algorithm, then image segmentation is 1 '\' ' , , ,

- L. , (UM; In UM; + CT, In CT, ) (7)

got . By topography, when a minimum point of two adjacent K
areas of the neighboring gray value is less than a given
threshold T, we combine them into one area instead of the The region growing threshold value is between [1, 50],
two water basin levees, as shown in figure Un figure 1, the figure 2 (a) is the graph of region growing segmentation, the
watershed of traditional watershed algorithm is ensured by formula (5) and (6) are the area of the image segmentation
dashed part, a watershed of the improved algorithm will be result, and their respective average, the result is shown in
figure 2 (b) and (c). we can see the tendency of increasing
ensured by solid part, so the over-segmentation will be
threshold segmentation regional internal uniformity from
avoided.
The water basin the diagram.

watershed

minimum value

(a) pepper
minimum value

Fig.1 Sketch of the improved algorithm OJ

For region growing algorithm of the optimal threshold T,

in this paper, based on image segmentation evaluation index
to determine a target function, by seeking the optimal value
of the objective function ,then the threshold is determined.

B. The Determination of the Objective Function

Recently there are many standards of the image 5 10 15 W � 30 35 � J5 50

[10.12]
segmentation , such as regional segmentation of image (b) contrast ratio
contrast, regional internal uniformity, they are classical and
commonly for image segmentation quality evaluation
099

criterion.
098

R, and Rj are two connected components, T is a 097

threshold region growing watershed segmentation, A, is the 096

size of R; area , J-LR, is the average gray of R, area, k; is the 095

094
number of R, ,we define the normalized coefficient below:
093

(3)
(c) homogeneity
Fig2 The change of uniformity and contrast
In order to get more accurately measure of
(4)
segmentation, we must ensure that the connected region of
segmentation algorithm is not too big and not too small, the
result of algorithm accords with the human visual
C r is the biggest variance for connecting component of characteristics, the internal region of the segmentation
image, and Dr is the biggest average difference between contrast and uniformity of extreme measure (maximum
the connecting part for adjacent grayscale. Write as:

52
 entropy and minimum entropy) are a part of the judging According to below formula:
standard, and form the objective function as follows: X/O)=[randx(B-A)/M +(B-A)/Mx(j-l)+A] (12)

� (Li ((UM; In UM; + CT,' In CT,')

Determine the position vector of the initial value of j a
H(T) = -
particles.
A large number of experiments show that: when T = 50,
+maxi UM; In max, UM; + min, UM; In mini UM; (8) under-segmentation is got, as shown in figure 3(e-h),on the
other hand ,over-segmentation is got when T is bigger than
+maxi CT,' In maxi CT,' + min, CT,' In min, CT,')) 50,as shown in

C. The Watershed Image Segmentation Algorithm Based on

Particle Swarm
Generally, the exhaustive method can get the maximum
value in the type (8). Because the time complexity of
exhaustive method is very high, it is not feasible in practical
application, The particle swarm optimization algorithm is
adopted to the optimal solution of formula (8), and the
optimal threshold image segmentation algorithm is the
parameter T.
13 14
Particle swarm optimization algorithm[ , ] is a kind of
optimization algorithm based on swarm intelligence and is
derived from the study of the flock foraging behavior. A
flock of birds search for food randomly, through the search
for a bird of the surrounding area, they change themselves
in the position and speed in the air, until they find food. (i) (j) (k) (I)
Particle swarm algorithm can solve the optimization Fig.3 Segmented results for images at different thresholds
problem and abstract as the optimal particle movement state. So the optimal value T is got between 10 to 50, and the
15
Clerc[ ]proposed shrinkage factor of particle swarm optimal value is solved between11 to 49, of course we can
optimization algorithm in 1999, particle swarm algorithm do adjust value T according to the different value.
has better convergence. In this paper, the optimal solution
When 10<T<50, in formula (11) we can take k = 0.729,
is got by the optimization algorithm of (8). The algorithm
needs to set the initial population size M and the speed of C1 = C2 = 2.05. We let M is 3, Vmax is 4, 1max is 40, the
the initial population. formula (12) is used to calculate the initial value, based on
the algorithm, this paper get the figure 3 (a), (b), (c), (d) the
V(O)=(V 1(0), ...,VM(O)) (9) convergence condition of the four images as shown in figure
And the position of beginning is shown below: 4 (a), (b), (c), (d). The optimal value is quickly converged in
house images and brain images of figure 4 in this paper, but
X(O)=(XI(O), ... , XM(O)) (10) pepper image, is slowly converged, Lena figure did not
Respectively ViCO) and XiCO) show the ith particle converge to the optimal value, the main reason for the house
velocity and position of beginning with brain images of the target relative to the single, and the
image does not exist in the grain area, and pepper image
k x[V,(I-l)+C, x ralldx (P'(I-l)-X,(I-I»
(11) target is relatively complex, Lena image texture exists in the
+C, x ralldx (G(I-l)-X,(I-l»] I v, (I) 1< v"" area.
V,(I) = -V=, V,(t)o;-V,_ -

V=, V,(t) 2: v,_

"

IV. THE RESULTS AND ANALYSTS OF EXPERIMENT "

In order to verify the effectiveness of the proposed

image segmentation algorithm, in this paper we analyze the
experimental results on Matlab 7.l.
160 S TO 15 20 25 JO 35 40 J5 '210 T 10 15 2Q 25 )0 )S

A. The Influence of The Parameter for the Result of the (a) lena Convergence (b) pepper Convergence
Experiments
The parameters of the algorithm in this paper is shown :
M is the population size, Vmax is the maximum speed of
particles, T range is growth parameters threshold, 1max is the
number of iterations and L is critical value.

53
" ,---"":;:::========:" [3] Ng H.P. Ong S.H. Foong K.W.C. et al. Masseter segmentation
using an improved watershed algorithm with unsupervised
"
classification [1]. Computers in Biology and Medicine,2008,38: 171-
"
1 84.
"
[41 Rao A.R, Srinivas V.V. Regionalization of watersheds by fuzzy
" cluster analysis [1]. Journal of Hydrology,2006,3 1 8: 57-79.
[5] Zhang M,Zhang L and Chen H.D. A neutrosophic approach to image
segmentation based on watershed method [1]. Signal Processing, 201 0,

: O�7-�,�O�"�����r-���"r-,�O�,,.
( C) house Convergence
t �O�'--�lO�"�����"r-,�,�"�,,.
(d) brain Convergence
[6]
90:1 510-1 51 7.
Flores F.C, Lotufo R.A. Watershed from propagated markers: An
interactive method to morphological object segmentation in image
Fig.4 Convergence for lena,pepper,house and brain images
sequences [1]. Image and Vision Computing, 2009,
doi:I0.1 01 6/j.imavis.2009.06.01 5.
B. Compared with Existing Algorithms
[7] Jung c.R. Combining wavelets and watersheds for robust multiscale
The image of figure 3(a) is segmented by the improved image segmentation [1]. Image and Vision Computing, 2007, 25: 24-
watershed algorithm. Figure5 (a) is the results of Lena image 33.
segmentation with improved watershed algorithm of [8] Wang X P. Morphological image analysis principle and
references [16], the Lena image is divided into 464 area, application[M]. Beijing: Tsinghua university press. 2008.

more traditional watershed and gradient had greatly [91 Wang D. A multiscale gradient algorithm for image segmentation
using watershed [1]. Pattern Recognition,1 997,30(12): 2043-2052.
improved watershed algorithm. Figure 5 (b) is the
[1 0] Zhang Y.J. A classification and comparison of evaluation techniques
segmentation results of Lena image in this paper, it is for image segmentation [1]. China Journal of Image and Graphics,
divided into 23 area, its results is better than traditional 1 996,1 (2):1 51 -1 58.
watershed algorithm, gradient watershed segmentation [II] Polak M, Zhang H, Pi M.H. An evaluation metric for image
[16] segmentation of multiple objects [1]. Image and Vision Computing,
algorithm and the algorithm mentioned in the literature
2009,27:1 223-1 227.
they have been improved greatly, and the segmentation result
[1 2] Cordenesa R, Luis-Garcia R and Bach-Cuadrab M. A
is consistent with the person's intuitive visual. Figure 5 (d) is
multidimensional segmentation evaluation for medical image data [1].
the segmentation results in this paper, the segmentation Computer Methods and Programs in Biomedicine, 2009, 96(2):108-
results of figure 5 (c-e) of the divider is superimposed on the 1 24.
original good segmentation results, figure 5(t) is the use of [1 3] Wang K.J, Guo Q.c. Image segmentation algorithm based on the
[16] PSO and improved Snake model [1]. CAAI Transactions on
literature the mentioned algorithm segmentation, it can be
Intelligent Systems, 2007, 2(1):53-58.
seen that this algorithm of segmentation results more in line
[1 4] Eberhart R.C, Kennedy 1. A new optimizer using particle swarm
with the intuitive visual. theory [1]. Sixth International Symposium on Micro Machine and
Human Science,1 995, (3):39-43.
[1 5] Clerc M. The swarm and queen: towards a deterministic and adaptive
particle swarm optimization [1]. IEEE International Congress on
Evolutionary Computation,1 999, (3): 1 951 -1 957.
[1 6] Luis Patino. Fuzzy relations applied to minimize over segmentation in
watershed algorithms[1]. Pattern Recognition Letters, 2005,(26):819-

�
828.

(d) (e)
>� (f)
Fig5 Comparison of results of different segmentation methods

ACKNOWLEDGMENT

This research is supported by the Project of Harbin

Normal University Research Fund (lIXY-03), Heilongjiang
Provincial University Engineering R&D Center of Machine
Vision and Intelligent Detection , Heilongjiang Provincial
Education Department Key Laboratory of Intelligent
Education and Information Engineering.

REFERENCES
[II Digabel H, Lantuejoul C. Iterative algorithms [1]. Proc 2nd European
Symp Quantitative Analysis of Microstructures in Material Science,
Biology and Medicine, Sturrgart West Germany: Riederer Verlag,
1 978: 85-99.
[2] Vincent L, Soille P. Watersheds in digital spaces: an efficient
algorithm based on immersion simulation [1]. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 1 991 ,1 3 (6): 583-598.

54