1. ABSTRACT
A novel method for image
processing and pattern recognition
using Discrete Fourier
Transformations on the global pulse
signal of a pulse-coupled neural
network (PCNN) is presented in
this paper.
The PCNN is an image
transform that removes
“unimportant” details while
improving the overall quality of the INTRODUCTION
During the last few years there was a
image. It may also provides shift of the emphasis in the artificial neural
substantial noise smoothing without network community toward spiking or pulse-
losing image pattern and shape coupled neural networks. Motivated by
biological discoveries, many studies consider
information.
pulse coupled neural networks with spike-
We describe the
timing as an essential component in
mathematical model of the PCNN information processing by the brain.
and an original way of analyzing Pulse-coupled neural networks
the pulse of the network in order to (PCNN) were introduced as a simple model
for the cortical neurons in the visual area of
achieve better quality of the
the cat's brain.These neural models are
image ,scale and translation-
proposed by Eckhorn and Johnson.
independent recognition for isolated . The essential model of PCNN, is
objects. described with details, that can be
implemented to perform a number of digital
image processing applications.
We describe in the next sections a
model that evaluates the global pulse of a
PCNN in order to find correlation in the pulse
signal and achieve pattern recognition.
IMAGE PROCESSING & its
Purpose:
1

2. Image processing is any form of information trained to fire (or not), for particular input
processing for which the input is an image, patterns. In the using mode, when a taught
such as photographs or frames of video; the input pattern is detected at the input, its
output of image processing can be either an associated output becomes the current
image or a set of characteristics or parameters output. If the input pattern does not belong in
related to the image. Most image-processing the taught list of input patterns, the firing rule
techniques involve treating the image as a is used to determine whether to fire or not.
two-dimensional signal and applying standard In PCNN,a neuron is operated in using mode.
signal-processing techniques to it. .
Basic purpose of image processing is for:
1. Improvement of pictorial
information for human
interpretation.
2. Processing of image data
for storage, transmission,
and representation for
autonomous machine
perception.
WHAT IS PCNN ? A simple neuron
A PCNN is a two-dimensional neural
network. They are treated as the third
WHY WE USE PCNN ??
In the field of digital image processing and
generation of NN models, that takes in to
pattern recognition , traditional models are
account spiking nature of neurons. Each
either subject to problems determined by
neuron in the processing layer is directly tied
geometric transforms (scaling, translation or
to an image pixel or a set of neighboring
rotation) or to high computational complexity.
image pixels, the two linking and feeding
Moreover, it is known today that parallel
inputs are iteratively processed and together
processing could solve determined by
to produce a pulse image with features, that
geometric transforms to take advantage of it
can be changed by varying the PCNN
we need parallelisable models.
parameters.
Neural models fits this requirement.
A Simple Neuron
A neuron is a device with many inputs and
IMAGE PROCESSING USING
PCNN
one output. The neuron has two modes of When PCNN is applied in image processing,
operation; the training mode and the using it is a single layer two dimensional array of
mode. In the training mode, the neuron can be laterally linked neurons.
2

3. STRUCTURE OF PCNN ♦ At each time step the neuron output .
♦ Y is set to 1 when the internal
♦ The number of neurons in the network activity U is greater than the threshold
is equal to the number of input image. One- function T. The threshold input at each time
to-one correspondence exists between image step is updated.
pixels and neurons. ♦ The output of the neuron is
consequently reset to zero when T is larger
♦ Each pixel is connected to a unique than U. Thus at one time step the pulse
neuron and each neuron is connected with the generator produces a single pulse at its
surrounding neurons with a radius of linking output whenever the value of U exceeds T.
field.
♦ The neuron receives input signals
from other neurons and ACTION OF PCNN IN IMAGE
from external sources through the receptive PROCESSING
fields. • Segment Ability
♦ After the receptive fields have Because of the local interconnections between
collected the inputs, they are divided into two the neurons, neurons encourage their
or more internal channels. One channel is the neighbours to fire only when they fire.
feeding input F and the other is the linking Thus, if a group of neurons is close to firing
input L. then one neuron can trigger the entire group.
♦ The feeding connections are required Thus, similar segments of
to have a slower characteristic response time the image fire in unison. This creates the
constant than those of the linking inputs. segmenting ability of the PCNN.
♦ The linking inputs are biased and
• Availability of Texture
then multiplied together, and further
information
multiplied with the feeding input to form the
The edges have differing neighboring activity
total internal activity U.
than do the interior of the object. Thus, the
♦ The pulse generator of the neuron
edges, which will still fire in unison, but will
consists of a stepfunction
do so at different times than do the interior
generator and a threshold signal signal
segments. Thus, the edges are may be
generator.
3

4. isolated. After several iterations the groupings This is all about the behavior of PCNN
of neurons tend to break in time. This “break- in image processing.
up” is dependent upon the texture within a What is Pattern Recognition ?
segment. This is caused by minor differences
Pattern recognition can be defined as "the act
that eventually propagate (in time) to alter the
of taking in raw data and taking an action
neural potentials. Thus, texture information
based on the category of the data".
becomes
available. A complete pattern recognition system
consists of a sensor that gathers the
• Denoising observations to be classified or described; a
For denoising, the intensity of a noisy pixel is feature extraction mechanism that computes
significantly different from the numeric or symbolic information from the
intensities of its surrounding pixels. observations; and a classification or
Therefore, most neurons corresponding to description scheme that does the actual job of
noisy pixels do not capture neighboring classifying or describing observations, relying
neurons or get captured by the neighboring on the extracted features.
neurons.
PATTERN RECOGNITION
USING PCNN
we can evaluates the global pulse of a
PCNN in order to find correlation in the pulse
signal and achieve pattern recognition.
• Smoothing PCNNs can be thought of as a combination of
Image smoothing is accomplished by two kinds of pattern recognition:
adjusting the intensity of each pixel based on • Statistical Pattern Recognition
the neuron-firing pattern in its neighborhood. In this kind,a set of features is extracted from
If a neuron fires sooner than a the pattern, grouped into a feature classes, and
majority of its neighbors fire, its intensity is recognition is based upon the partitioning of
adjusted downwards. the feature space in such a manner that new
If a neuron fires after the majority of views are classified properly.
its neighbors have fired, its intensity is
• syntactic pattern recognition
adjusted upwards.
It deals with the relationship between the
If a neuron fires with the majority of
features as well as the features themselves, so
its neighbors no change is made. After
that the face in that partially damaged picture
completing the firing
would still be recognizable.
cycle (all neurons fire exactly once) the
network may be reset by forcing the threshold
The problem with statistical recognition is its
values of all neurons to zero and the
partially limited processing ability, and the
smoothing process may be repeated.
problem with syntactic lies in the difficulty of
gaining real time information (it's very slow).
4

5. However, when the two are combined, the Information flow is mainly feed-forward but
end result is a system that comes extremely there are also lateral interactions between the
close to obtaining the level of pattern pulse-coupled neurons.
recognition that humans possess.
ARCHITECTURE OF
• The Pulse Coupled Neural
MODEL
Networks
The model proposed here is based on three
The key of the entire system lies in
modules of processing:
the neural analyzer that, in our case, is made
• The pulse coupled neural of pulsecoupled neurons, which act like local
network analyzer cells .
• The Discrete Fourier
Transform (DFT) module
• The multilayer perceptron
(MLP) classifier.
♣ The pulse train generated by the
neurons is a direct result of stimulus
5

6. excitation and lateral interaction between Sijkl is the stimulus component
neurons. computed from the pixel intensity (<i+k, j+l>,
♣ Lateral interaction and further "<x,y>" meaning the intensity of the pixel
stimulation determine the neurons to fire in with coordinates x and y) in the input
synchrony in the homogenous areas image.Usually this value is normalized.
associated to the image. These effects can be ♣ VF and VL are normalizing
exploited in image segmentation. However, constants and M and W represent the constant
our assumption is that the pulse train of the synaptic weights. M and W are computed by
neurons captures somehow morphological using the inverse square rule
information from the image neurons captures f(k, l)=2/√ ( k2 + l2 )
somehow morphological information from Y stands for the output of the neuron and can
the image. only take a binary value of 0 or 1.
In the next equations we will refer
♣ The linking effect can be modeled
to “n” as being the current iteration (discrete
as follows:
time step) where "n" varies from 1 to N-1 (N
N is the total number of iterations; n = 0 is Uij [n] = F ij [n] .(1 - β. L ij [n] )
the initial state).The dendritic tree can -(3)
bedescribed by Thedendritic tree can be Uij[n] represents the internal activation of the
described by e neuron and β is the linking weight parameter.
following equations: ♣ The pulse generator determines the
-αF
Fij[n] = e .Fij [n - 1]+ firing events in the model. In fact, the pulse
 V F.∑ kl M kl S ijkl --(1) generator is also responsible for the modeling
of the refractory period.
Lij [n] = e-αF. Lij [n - 1] +
♣ As the neuron produces a spike, its
V L.∑ kl W kl Yijkl[n-1] --(2)
threshold is raised to prevent it from firing
The two main components F and L again in the near future (established by the
are called feeding and linking. The
(i,j) pair stands for the position of the neuron
in the map. αF and αL are time constants for
feed and link.
6

7. ♣ For each iteration the total number
parameter settings). The threshold is then of firings (Equation 6) over the entire PCNN
decreased to allow the neuron to fire when its is computed and stored in a global array G
activation is increased. (see Fig. 1).
G[n] =∑ ijY ij[n] , --(6)
1,if Uij[n] >Θij [n - 1]—
where n is the iteration (n =0 …. N-1)
(4)
♣ The global array is then used at the
Y ij[n] ={ 0,otherwise next levels of the system (to compute the DFT
of the global pulse signal).
Θij [n] = e-αΘ .Θij [n - 1]+VΘ.Y ij –
(5) • The Discrete Fourier
In equations (4) and (5) Θ ij[n] Transformations
represents the dynamic threshold of the
We used the standard analysis
neuron while αΘand VΘ are the time constant
equations to calculate the DFT:
and the normalization constant respectively.
♣ During the simulation, each Re X[k]= ∑N-1
i=0 G (i ) cos(2Πki
iteration updates the internal activity and the
--( 7 )
output for every neuron in the network, based /N) k=0,…N/2
on the stimulus signal from the image and the
previous state of the network.
7

8. lm X[k]= ∑ N-1
i=0 G (i ) sin(2Πki
 PCNN s can be used in several tasks
of pattern recognition such as Face
-( 8 ) recognition,Finger print
/N) k=0,…N/2
identification etc.
Computing the DFT means basically
 PCNN s can be used in Image to
correlating the input signal with each basis
sound converters where the data
function.
produced by the PCNN (usually in
♥ The DFT yields two shorter signals
the form of icons that can be
to be analyzed. We used only the imaginary
represented by only a few bits) will
part of the DFT in further processing but a
be gathered and encode it for the
combination may be possible as well.
sound generator.
Our choicehad been motivated by
experimental observations that show a
relative stability of the real part over all the
shapes used for testing.
♥ We also enhanced speed by using
only the imaginary part in the higher levels.
CONCLUSIONS
• The classifier Single-layered, laterally linked pulse-coupled
Our classifier is basically a multilayer neural networks, being fairly insensitive to
perceptron (MLP). The neural architecture local intensity variations and noise in digital
consists of one input layer, one hidden layer images, are highly effective for image
and one output neuron. smoothing without
♦ The input layer contains a number of blurring and eroding or dilating edges.
inputs equal to the samples in the imaginary Furthermore, PCNNs are able to
part of the DFT signal (Im X in eq. (8)). perform image segmentation.
♦ Then, a hidden layer has an extension The PCNNs are presented in different models,
not all the models are fully evaluated to
of about 10 to 20% of the input layer.Because
discover its potentials in image process.
of the specific tasks used to test the system.
However, this paper presented an evaluation
♦ The output layer contained only one
for one of the PCNNs models, against many
neuron (target detection). An output value of
variables that control the output..
1 is equivalent to target detection whereas a
Since PCNNs are still a rather
value of 0 means no target detect.
recent development, their future will
undoubtedly bring a wealth of
APPLICATIONS OF
opportunities and challenges alike. Among
PCNN the current problems that researchers are
 PCNN s can be used in several tasks actively working on
of image processing, such as image is:
segmentation, edge extraction, object  An exact determination of the
identification, object isolation. relationship between the various
8

9. parameter values and the
performance of the network.
 The establishment of an automated
PCNN system, which can set
parameters optimally.
 Designing digital and optical
hardware for the PCNNs, which can
function in real- time applications.
REFERENCES
1) Image processing with pulse-coupled
Neural Networks --J. M. Kinser, T.
Lindbladh, -- Springer- Verlag London
Limited,
2)Image Processing with Neural Networks-a
review -M. Egmont-Petersena., D. de
Ridderb, H. Handelsc,
3)Network of spiking Neuron -The Third
generation of Neural Network model ,Neural
Network--W. Maass,
4)www.sciencedaily.com/releases
5)www.yet2.com
9