880 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 22, NO. 6, JUNE 2011
A New Automatic Parameter Setting Method of a
Simplified PCNN for Image Segmentation
Yuli Chen, Sung-Kee Park, Yide Ma, and Rajeshkanna Ala
Abstract— An automatic parameter setting method of a simpli-
fied pulse coupled neural network (SPCNN) is proposed here. Our
method successfully determines all the adjustable parameters in
SPCNN and does not need any training and trials as required by
previous methods. In order to achieve this goal, we try to derive
the general formulae of dynamic threshold and internal activity of
the SPCNN according to the dynamic properties of neurons, and
then deduce the sub-intensity range expression of each segment
based on the general formulae. Besides, we extract information
from an input image, such as the standard deviation and the
optimal histogram threshold of the image, and attempt to build a
direct relation between the dynamic properties of neurons and the
static properties of each input image. Finally, the experimental
segmentation results of the gray natural images from the Berkeley
Segmentation Dataset, rather than synthetic images, prove the
validity and efficiency of our proposed automatic parameter
setting method of SPCNN.
Index Terms— Automatic parameter setting, dynamic property,
general formulae, image segmentation, optimal histogram thresh-
old, simplified pulse coupled neural network, standard deviation,
static property, sub-intensity range.
I. I NTRODUCTION
E
CHORN’S cortical model, a bio-inspired neural network,
was developed in light of synchronous dynamics of
neuronal activity in cat visual cortex [1]–[3]. According to
its ability to cause the adjacent neurons with similar inputs
to pulse synchronously, it was soon recognized as having
significant potential in image processing. Therefore, pulse cou-
pled neural network (PCNN) model, a tailored and modified
version of Echorn’s cortical model, was developed by Johnson
et al. and became more suitable for image processing [4]–[10].
Nowadays the standard PCNN model is usually simplified to
achieve lower computational complexity while remaining the
essential visual cortical property. The representative simplified
Manuscript received March 20, 2010; revised March 2, 2011; accepted
March 6, 2011. Date of publication May 5, 2011; date of current version
June 2, 2011. This work was supported in part by the National Science
Foundation of China under Grant 60872109 and the Korea Institute of Science
and Technology.
Y. Chen is with the School of Information Science and Engineering,
Lanzhou University, Lanzhou, Gansu 730000, China. She is also with the
Center for Cognitive Robotics Research, Robotics/Systems Division, Korea
Institute of Science and Technology, Seoul 136-791, Korea (Corresponding
author e-mail: chenyuli2008@live.cn).
S.-K. Park and R. Ala are with the Center for Cognitive Robotics Research,
Robotics/Systems Division, Korea Institute of Science and Technology, Seoul
136-791, Korea (e-mail: skee@kist.re.kr; ala_rajeshkanna@yahoo.co.in).
Y. Ma is with the School of Information Science and Engineering, Lanzhou
University, Gansu 730000, China (e-mail: ydma@lzu.edu.cn).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TNN.2011.2128880
models include the intersecting cortical model [11]–[13], the
unit-linking PCNN model [14], [15] and the spiking cortical
model (SCM) [16].
The past decade has seen the rapid development of PCNN
in many aspects within the image processing field, such as
image segmentation [7], [17]–[24], image shadow removal
[25], image understanding [26], feature extraction [5], [14],
[27], [28], pattern recognition [29], [30] and object recognition
[31]–[33]. More detailed applications of PCNN can be found
from literatures of Lindblad and Kinser [13], Ma [34] and
Wang [35].
Among the applications in image processing, PCNN has
great potential for developing image segmentation algorithm.
However, the quality of the segmentation results strongly
depends on the appropriate values of PCNN parameters.
Moreover, the appropriate values of the parameters could only
be adjusted manually or estimated through a heavy training,
which has become a serious problem and constrained the
further development of PCNN.
Many researchers have attempted to solve this problem. For
example, Kuntimad and Ranganath [17] determined the mini-
mum amplitude value of dynamic threshold which guarantees
each neuron fires exactly once during a pulsing cycle, and ana-
lyzed the minimum and maximum values of linking strength β
for a perfect image segmentation after acquiring the intensity
ranges of object and background. Based on this principle of
linking strength β , Karvonen [18] computed a specific value of
the linking strength β according to the distribution information
of the segments obtained from training synthetic aperture radar
images. Stewart et al. [19] proposed a region growing PCNN
for multi-value image segmentation, but it requires to manually
set the time-variant dynamic threshold E and linking strength
β with incremental constants. The method introduced by Bi et
al. [24] determined the weight matrixes and linking strength
β adaptively according to spatial and gray characteristics, but
they empirically set the amplitude and decrement constant
of dynamic threshold E . In addition, Yonekawa et al. [23]
automatically adjusted the linking strength β , synaptic weight
W and exponential decay coefficient α
e
but it required many
iterations of trials.
So far, most of the previous methods could not set all
the parameters of PCNN automatically, except the methods
proposed by Berg et al. [21] and Ma et al. [22]. The former
is to evolve PCNN neurons by automatic design of algorithms
through evolution, while the latter is to build an automated
PCNN system based on genetic algorithm [22]. However,
these algorithms require repetition of trials and a heavy
1045–9227/$26.00 © 2011 IEEE