Neural Network Based Automatic Traffic Signs Recognition

Abstract

Image recognition and understanding is one of the most interesting fields of researches. Its main idea is to bridge the gap between the high level human image understanding and the low level machine image representation. Quite a lot of applications have been suggested in different fields like medicine, industry, robotics, satellite imagery and other applications. This paper proposes a new approach of traffic signs image recognition and understanding using computational intelligent techniques and the application of this approach on intelligent cars which can recognize the traffic signs and take a decision according to the signs it reads. Supervised machine learning has been selected since the algorithm does not need to classify the images but to identify their precise meaning. Different neural networks have been trained and used in this paper. The best neural network has been selected, which uses genetic algorithms in its training, and is known as evolutionary training neural network. Different image features have also been investigated and discussed. the best of these features, which fit the requirement of the suggested algorithm, have been selected.

Full text

Neural Network Based Automatic Traffic Signs Recognition
Mohammad A. N. Al-Azawi
Oman College of Management and Technology
P. O. Box: 680; Postal Code: 320; Barka; Sultanate of Oman
mohd.alazawi@omancollege.edu.om
ABSTRACT
Image recognition and understanding is one
of the most interesting fields of researches.
Its main idea is to bridge the gap between
the high level human image understanding
and the low level machine image
representation. Quite a lot of applications
have been suggested in different fields like
medicine, industry, robotics, satellite
imagery and other applications. This paper
proposes a new approach of traffic signs
image recognition and understanding using
computational intelligent techniques and the
application of this approach on intelligent
cars which can recognize the traffic signs
and take a decision according to the signs it
reads. Supervised machine learning has been
selected since the algorithm does not need to
classify the images but to identify their
precise meaning. Different neural networks
have been trained and used in this paper.
The best neural network has been selected,
which uses genetic algorithms in its training,
and is known as evolutionary training neural
network. Different image features have also
been investigated and discussed. the best of
these features, which fit the requirement of
the suggested algorithm, have been selected.
KEYWORDS
Image recognition, Neural Nets,
Evolutionary Training, image features,
Supervised Learning,
1 INTRODUCTION
Due to the tremendous increase in the
use of images in different application,
and the advance in technology, image
recognition and understanding has
become one of the most attractive fields
of research. The main purpose of image
understanding and recognition is to
bridge the gap between high level
understanding human, and machine low
level representation of images. Humans
can see images, interpret and understand
them depending on the knowledge
acquired from their accumulative
learning process, while computer can
represent images in terms of zeros and
ones. The machine image recognition is
widely used in different fields such as;
robotics, security, industrial, medicine,
etc.
Features can be defined as measures that
can be extracted from the image to be
used in image understanding. Different
features can be extracted from the
image, like colour, texture, and shape.
These features are used to identify the
contents of the given image. This
identification process is known as
Content-Based Image Retrieval (CBIR).
CBIR was used for the first time in
1992 by T. Kato to describe processes
of automatic retrieval of images from
a database [1]. CBIR utilizes computer
vision principle in describing the
contents of an image, rather than
describing the image as a whole. In
this paper, the principle of CBIR will
be utilized in additional to
computation intelligence to develop
an algorithm that can be used by
machines (cars in this case) to
recognize traffic signs and identify
their meaning. Decisions and
appropriate actions can be taken
according to the traffic sign meaning.
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
753

Computation intelligence and machine
learning is used in this algorithm to
improve and speed up the recognition
process. Neural network, which is
widely used in machine learning, is
selected to be used due to its ability of
learning. The supervised learning
technique has been selected to be
used since the machine needs to
identify the precise meaning of the
sign to take actions accordingly.
2. IMAGE PROCESSING,
RECOGNITION, AND
UNDERSTANDING
Pictures, also referred to as images, as
commonly used in the field of
computer technology, can be defined
as any visual representation of scenes,
persons, or any other object.
Computers can see an image as a set of
pixels or in more specific words, as a
set of zeros or ones. To make the
image useful in different applications,
it is important to analyze this image
and describe it automatically, using
computer image processing and
computer vision.
Many models of image representation
are used to represent images in digital
formats like RGB, HSL, VSL, and other
models [2]. The most widely used model
in image processing is the gray-level
image representation model. In this
model, the image
is represented
by the luminance
of the pixels at the
spatial location
and . Figure 1
shows the gray level representation of an
image, in which the luminance value is
dependent on the number of bits which
are used to represent the pixel [3].
The number of gray levels (luminance
levels) L is dependent upon the number
of bits used to represent the pixel value,
as given below:
1
Where L is the number of gray levels
and n is the number of bits.
8 bits (1 byte) representation is widely
used, which gives 256 gray levels.
Figure 1 Gray image representation
2.1 Image Processing
Image processing consists of algorithms
that are applied on the pixels values of
the image to perform a specific process,
such as enhancement, restoration, feature
extraction, etc.
As shown in Figure 2, the input raw
images go through the processing
algorithms to produce a processed
image. The resulting image might be
visually enhanced, filtered, etc. In case
of features extractions algorithms, some
feature vectors can be extracted from the
image. The general formula for an image
processing algorithm is given by:
2
Where
is the processed image
resulted from applying the processing
on the original image
.
120
125
130
70
99
123
190
143
98
98
90
188
180
82
66
72
78
67
87
64
(x,y)
x
y
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
754

Figure 2 Image processing system
2.2 Image Features Extraction
In order to describe the image contents
to make the understanding process easy,
features should be extracted from the
image. Many features have been
suggested and extracted from the image
by applying some image processing
techniques. There are mainly two types
of features; low level and high level
features. Low level features can be
extracted directly from the image, like
edges, lines, textures, etc. The low level
features are not easy to be used in image
understanding since they do not provide
information about the relationship of
shapes in the image. Therefore, a set of
high level features can be extracted from
low level features, which are useful in
image understanding and content
recognition.
1) Edge Map
Edges which are widely used in different
applications, can be extracted by
applying a high pass filter on the image.
High-pass filters like Gaussian and
Laplacian are used to extract the edges,
since the variation of the luminance at
the edges is high. In other words, first
and second derivatives can be used in
edge extraction.
First Order Edge Detection
For an Image
the vertical and
horizontal derivative can be obtained as
follows:
3
4
The magnitude and angle of the edge is
found as follows:
5
6
The above equations assume that the
image is a continuous function, while in
computer representation, the image is
always discrete, thus the above equations
can be represented as:
7
8
Second Order Edge Detection
The second derivative is greater when
the change of the first derivative is
greater, and zero if it is constant. The
second derivative equals zero when
the first derivative is at the maximum
point, which is constant. Laplacian
Second Order Operator is one of the
commonly used techniques in
extracted edges from an image. For
the image
in the continuous
domain, the Laplacian can be found
from the following equation:
9
10
For a discrete image, the Laplacian can
be found as follows:
11
2) Colour Features
Colour is another important feature that
can be used to describe the nature of the
image, depending on the distribution of
the colours. Histogram is the most
Raw
Image
Image Processing
Algorithms
Processed
Image
Features
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
755

commonly used colour feature
descriptor. Image colour histogram is
the distribution of colours over image
pixels, or simply the frequency of
occurrence of colors in the image. The
main advantage of using color
histogram is that it is not affected by
the three main camera changes, which
are rotation, scaling, and shifting. The
biggest weakness of using histogram
is that it loses the space information of
the color [4].
Color histogram
, is defined as a
vector as follows:
12
Where
is number of pixels of color
level j in the image and
is the
number of colours or gray levels in an
image.
Color Moments, usually, the mean
(first order), variance (second order),
and skewness (third order) are used
as the features which are used to
describe the histogram and the colour
contents of the image eventually. The
moments are given below:
13
14
15
Where
is the value of
color
component of the image pixel
, and
is the number of pixels in the image.
3) Texture Features
Although texture is not well-defined,
like color feature, it gives a good
description of the content of the object in
the image, like cloud, trees, bricks, and
fabric. Texture is considered as a high
level semantic for the image retrieval
processes.
Texture features can be obtained using
Gabor filter, wavelet transform, co-
occurrence matrices, and local statistics
measures. Among the six Tamura
features; coarseness, directionality,
regularity, contrast, line-likeness,
contrast and roughness, the first three are
more significant. The other three are
related to the first three and do not add
much to the effectiveness of texture
description [5].
3. IMAGE SEGMENTATION
Segmentation is the process of dividing
the image into non-overlapping,
homogenous, and connected regions [6].
The image is segmented according to
some measurements and features like
gray-level, colour, texture, edges, etc.
Reference [7] introduces a good survey
of segmentation techniques. Another
important survey was published in 2011
by Thilagamani and Shanthi [8], who
have published a survey on segmentation
through clustering. They have presented
different techniques in segmentation,
and they defined the clustering as the
grouping of similar images in the
database. Adaptive clustering techniques
and generalized K-Mean algorithm in
clustering were used by Pappas [9].
Artificial intelligence-based
segmentation was used by many authors.
Deshmukh and Shinde who have used
neuro-fuzzy system for color image
segmentation [10].
Segmentation algorithms may be
classified furthermore as local and
global segmentation algorithms. In local
segmentation, only the features of the
pixels values and their neighboring
pixels are considered, while in global
segmentation the image as a whole is
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
756

considered in the segmentation process
[11].
Segmentation divides the image
into a
set of homogenous, non-overlapped,
connected subregions
. The union of
these subregions should form the
original image.
16
The union of all subregions forms the
original image
17
Each regions
should be homogeneous
for all i = 1, 2, …N.
Finally different adjacent regions
and
should be disjoint.
18
Colour-Based Image Segmentation
Colour is an important feature that
recognizes the regions in an image.
Based on a human’s vision system
(HVS), the colours of an image are
reduced to a specific number of colours
which is known as the colour
quantization process. In a human’s
vision system, humans recognize only a
few colours, and he/she does not need to
know the precise colour bands values,
while in the machine vision system
(MVS) the machine can recognize
millions of colours. Based on the above
reality, one can divide the image
according to the colours of the regions;
e.g. the sky is blue in the HCS point of
view, where there are few thousands of
colours in MVS. Referring to the above
example of the sky, it is easier to extract
the region which represents the sky if it
is defined as blue regardless of the
difference in blue brightness or darkness
level.
Many techniques in colour quantization
have been used till now using different
colour model representations, like RGB,
XYZ, HSV, etc.
The RGB colour model is the simplest
and most important one. A colour
lookup table is used to map colours into
a certain colour class. Some standard
colours are used in the table and all other
colours are mapped to these colours. The
distance between every colour in the
image and the standard colours is
calculated. The standard colour with
minimum distance is selected to replace
the colour under process.
Figure 3 shows the RGB colour space in
which the coordinates of any colour
point can be described using three
coordinates (r, g, b). Standard colours
can be extracted from the vertices of the
cube. Pure red can be described as the
point with the highest r components
value (1), and both green and blue values
are 0 i.e. (r, g, b) = (1, 0, 0). In the same
way different standard colours can be
extracted as shown in Table 1. The total
number of colours in the lookup time
table can be found as follows:
19
Where:
is the total number of
colours in the lookup table,
is the
number of components used to represent
each colour, and
is the total number of
values can be used for each component.
Since the number of components in each
colour is 3 and there are 2 values for
each component
then the total
number of standard colours in the lookup
table is
.
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
757

Figure 3 RGB colour model space cube
Table 1 Standard RGB colours lookup table
R
G
B
Colour
Short
0
0
0
Black
L
0
0
1
Blue
B
0
1
0
Green
G
0
1
1
Cyan
C
1
0
0
Red
R
1
0
1
Magenta
M
1
1
0
Yellow
Y
1
1
1
White
W
The distance between any point in the
colour space P, which has the
coordinates (x, y, z) and the standard
colour that is given in the lookup table
(r, g, b) is calculated using the formula
given in Equation 20. The point takes the
nearest standard colour, i.e., the standard
colour with minimum
.
20
4. COMPUTATIONAL
INTELLIGENCE IN IMAGE
PROCESSING
Artificial Intelligence (AI) can be
defined as “simulation of human
intelligence on a machine, in order to
make the machine efficient to identify
and use the right piece of knowledge at a
given step of solving a problem” [12].
The main goal of AI is to add the ability
of learning and thinking to computer
applications. AI has many applications
in different fields like Expert Systems,
Speech and Natural Language
Understanding, Intelligent Control,
Image Understanding, Computer Vision
and many other applications [13].
Production systems, Swarm Intelligence,
Neural Nets and Genetic Algorithms are
examples of the techniques that are
widely used in AI applications.
Machine learning is the most important
part of AI, in which the machine can
adapt its knowledge according to the
input it gets from the environment. One
of the important applications of AI
algorithms is in the field of image
processing applications, especially in the
field of image recognition and computer
vision. Since the early nineties, AI was
widely applicable in image processing
applications, both in low level and high
level processing. Low level image
processing represents applications like
noise removal and image enhancement
and restoration. High processing
represents applications like semantic
features extraction, as well as image
recognition and understanding, or in
general computer vision applications.
Computer vision and understanding is
the process of identifying the image
contents and search for similar contents
in other images. This process requires
some kind of intelligence and the
learning ability. The conventional search
techniques are suffering from a serious
problem. They become very slow as the
image database gets larger because the
algorithms need to search all the items in
the database. In order to find similar
images faster, many algorithms have
R
G
B
(0, 1, 0)
Green
(1, 0, 0)
Red
(0, 0, 1)
Blue
(1, 1, 1)
White
(0, 0, 0)
Black
(1, 1, 0)
Yellow
(0, 1, 1)
Cyan
(1, 0, 1)
Magenta
(r, g, b) P
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
758

been proposed to reduce the search time.
AI is one of the efficient techniques that
have been used in Computer Vision to
reduce the search time and to improve
the performance since it gives machines
the ability to learn and think.
4.1 Machine Learning
Machine learning utilizes the available
information to improve machine
understanding, and then increases their
performance.
In other words, how to give the machine
the ability to learn in once problem to
solve similar problems. Mainly there are
two types of machine learning,
supervised and unsupervised.
Nevertheless there is a third learning
technique which is known as
reinforcement learning.
1) Supervised Learning
In this type of learning, the machine is
provided with input and desired output.
The training process is improved by
finding the error ratio of the actual
output and the desired output. The
machine gathers some knowledge from
the training process in such a way that it
can give correct responses to similar
inputs. Inductive learning and analogical
learning are the most well-known
techniques that are used in supervised
learning. Many other techniques were
proposed in training the neural networks.
2) Unsupervised Learning
If the desired output is not known,
unsupervised learning is used. In this
learning category, the machine is
provided with input only, and it should
update itself to generate classes for
similar objects or the objects with
similar features.
3) Reinforcement Learning
This type of learning is intermediary
form of supervised and unsupervised
learning. The learning machine performs
action on the environment and gets a
feedback response from the
environment. The learning system
grades its action good or bad based on
the environmental response, and
accordingly adjusts its parameters [12].
4.2 Neural Networks
Neural Computing is widely used in
computer vision applications, since it
offers the possibility of adaptation and
learning. In such algorithms the input to
the Neural Net (NN) is a vector of
features that are extracted from images.
This vector of features is used to match
the contents of an image with other
features vectors stored in a database.
Supervised and non- supervised learning
is used in retrieving images from the
database. In most proposed systems,
supervised learning is used because it
gives the possibility to improve the
retrieval performance. Such retrieval
systems utilize user feedback as part of
the feedback to supervised learning NN.
Laaksonen et al. (2001) described the
implementation of relevance feedback
by using Self-Organizing Maps [14].
Hybrid Neural Networks were used in
image classification which was used in
image retrieval process by Tsai et al. in
2003 [15]. Relevance feedback has been
widely used in Neural Network – based
image retrieval systems like in [14],
[16], [17], [18] and [19]. Many systems
have utilized the Radial Basis Function
Network to improve the performance of
Neural Networks in retrieving images
[20], [19]. Lee & Yoo, (2001) have
introduced a Neural Network – Based
computer vision techniques system and a
Human Computer Interaction approach
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
759

to CBIR using the Radial Basis Function
(RBF) network [20]. In 2011,
Nematipour et al. propposed what they
called Enhanced Radial Basis Function
Network and Relivance Feedback to
design an effective image retrieval
mechanism in CBIR [19]. Liu and Mio
proposed in 2007 an algorithm that uses
spectral histogram features to describe
the spatial relationship among pixels as
input to their neural network to organize
images for CBIR [21]. The Neuro-Fuzzy
approach was used with the 2-D wavelet
Transform by Balamurugan and
Anandhakumar in 2009 as a tool for
clustering images in computer vision and
CBIR tachniques [22]. Cubic – Splines
Neural Networks have been used to
reduce the gap between high level
concepts and low level visual features
that are used in image retrieval [23] [24].
The application of multilayer neural
network was suggested by Rao et al. in
2010 for automatic image retrieval [25].
Characteristics of ANN:
1. Nonlinearity: means that the NN
consists of a huge number of
interconnections between nonlinear
neurons.
2. Input/output mapping: mapping means
that an ANN gives output for each
input; this mapping process consists of
some learning activities.
3. Adaptability: means that NN can adapt
its free parameters to changes to the
surrounding environment
Learning Mechanisms of Neural Nets
In order to get reasonable results, neural
nets should be trained first. Learning or
stimulation can be defined as the process
of changing the free parameters of neural
nets, this is done by adapting the
connections weight between the neurons
or synaptic. The learning process is
iterative and can be classified mainly
into supervised and unsupervised
learning techniques. Many techniques
have been proposed to achieve this:
1. Error Correction Learning
2. Memory – Based Learning
3. Hebbian Learning
4. Competitive Learning
5. Boltzmann Learning.
What is important in this paper is the
Error Correction Supervised Learning
Technique, which will be used to train
the neural network.
6. PROPOSED ALGORITHM
Figure 4 shows the flowchart for the
proposed algorithm, in which the image
is acquired by a camera and sent to the
first stage which includes segmentation
algorithms. In this paper RGB-Cube
colour based segmentation is used.
The segments are going through a
feature extraction algorithm in which
some features are extracted to be used in
the recognition process. Different
features have been used such as; colour,
edges, texture, and binary images.
A knowledge database is used to train
the neural network before starting the
recognition process. The neural network
is used to recognize the meaning of the
traffic sign. Different neural networks
have been studied, back propagation NN
has been selected in the training and
recognition process. For speeding up the
training process Genetic Algorithm has
been used to update the weights of the
NN, which is known as evolutionary
back propagation training technique.
1) Image Acquisition
The image is acquired using a regular
camera with a frame rate of at least one
frame per second. Most fixed pictures
cameras have frame rate greater than two
frames per second. Video cameras can
be used as well. Video cameras have
faster frame rate, but less resolution. The
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
760

images will be processed frame by
frame. Let’s assume that a car is moving
in the fastest allowed speed, which is
120 kilometers per hour. This means that
the car is moving
meters
per second. With a camera of 2 frames
per second, it is possible to take a shot
every 12 meters approximately. The time
period between each two shots is
sufficient to process the image and get
the result, since there is no image search
required and the processing time is used
to segment the image, extract features,
and identify its contents by NN.
Figure 4. Recognition Flowchart
2) Segmentation Process
The input image is segmented into
regions, one of the regions contains the
traffic sign which will be identified and
recognized. Figure 5 shows the results
obtained from applying the
segmentations algorithm. The original
image is undergoing a colour reduction
process in which the number of colours
is reduced to only 8.
(a)
(b)
(c)
(d)
(e)
(f)
(a)
(b)
(c)
(d)
(e)
(f)
Figure 5. Image segmentation process, (a)
original image, (b) colours-reduced image, (c),
(d), (e), and (f) image segments (regions).
The resultant image is shown in (b). The
reduced colour image is segmented
according to the colour and regions, the
images given in (c), (d), (e), and (f)
shows the extracted regions. One of
these regions contains the sign which
needs to be recognized. From (f) it is
clear that the background has been
cleared as much as possible.
3) Features Extraction
All regions except the one that contains
the sign are neglected and the sign is
processed more to extract some features
that can be used as input to the neural
net to be identified. Many features have
been used, but only one has been
selected. Colour, shape, and texture can
be used in this process. The colour
feature is not suitable in this application
Acquired
Image
Segmentation
Algorithm
Extract Feature
for Each segment
Knowledge
DB
Train NN
Trained NN
Recognized
Traffic Sign
Take an Action
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
761

since most of the traffic signs may have
the same colour distribution. Texture is
not suitable also since most traffic signs
are smooth (mostly smooth red, smooth
blue, etc.).
(a)
(b)
(c)
(d)
(e)
(f)
Figure 6. Feature extraction; (a) original image,
(b) edge map, (c) binary image using Bradley
local, (d) binary image using iterative
thresholding, (e) fuzzy bimodal thresholding
(dark) (f) fuzzy bimodal thresholding (light).
Edge and binary image features are used
as input to the neural network. The
image is reduced in size in such a way
that it keeps the necessary information in
the image and reduces the input size of
the neural network. Edge features did
not give reasonable results since the
edge map do not give clear features as
shown in Figure 6 (a).
To generate the binary image, many
thresholding techniques have been tried,
such as Bradley local, iterative, global,
local, and fuzzy-based bimodal
thresholding (FBB) [3].
As shown in Figure 6 (c) through (f) the
best results were obtained in iterative
and FBB. Thus FBB has been adopted to
generate the query image that will be
used as an input to neural network for
recognition.
4) Traffic Sign Image Recognition
The resultant image will pass through
the recognition process, in which the
image is entered into a trained neural
network and the decision is made
according to the result of the NN.
The neural network needs to be trained
first, using a set of sign images. The
training may take some time, from few
minutes to a couple of hours. This
training is needed only once, unless new
data is used.
Supervised learning was used since the
algorithm needs to learn about the
meaning of each sign to take an
appropriate action according to that
meaning.
Back propagation (BP) and delta rule
(DR) training techniques have been
examined. Back propagation is slower
than Delta rule in training but it gives
better results due to the nonlinearity
property. Nonlinearity is a result of the
hidden layer the BPNN consists of.
7. RESULTS
A total of 40 samples were used to train
the neural network. The images used are
of size of
,
, and
pixels. Both and
3
have given similar results while
has taken a longer time in
training the NN, thus 3
has been
used, which gives a total of 900 input
nodes. The neural network was designed
to contain one hidden layer with 16
neurons in the hidden layer and the
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
762

number of the outputs is equal to the
types of signs to be recognized.
Figure 7 shows the effect of the number
of neurons in the hidden layer on the
training and recognition processes. It is
clear from the graph that as the number
of neurons increased the training time is
increased as well as the number of
required iterations. While the accuracy is
not increased in a high rate before the
scale 16, at which the accuracy increased
became almost constant. This will help
in making decision about the optimized
number of neurons in the hidden layer
which is in this case was 16.
8. CONCLUSIONS
As a conclusion to the above discussion
one could note that the algorithm is
applicable on the specified application
since it is very fast and does not need
any real time search in the database.
Neural net gave strength to this
algorithm since it does not need a long
time to recognize the image. The main
time will be consumed in the processing
of the image, which is in microseconds.
Back propagation training has oscillated
in some training cases, and became
unstable. This problem was solved using
revolutionary back propagation (EPB)
technique, in which the genetic
algorithm was used to find the optimum
weight for the neurons. In some cases
(EPB) took longer than PBNN to be
trained, but the main advantage of EPB
is that it does not oscillate, and it
converges to a stable NN in all the
training processes in this paper.
(a)
(b)
(c)
Figure 7. The effect of number of neurons in
the hidden layer on training and recognition
statistics; (a) number of iterations, (b) training
time, and (c) recognition accuracy.
9. REFERENCES
1
S. Kulkarni, " Content-Based Image
Retrieval," Vishveshwaraiah Technological
University, 2009 - 2010.
2
"Wikipedia / HSL and HSV," [Online].
Available:
http://en.wikipedia.org/wiki/HSL_and_HSV.
[Accessed 1 9 2011].
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
763

3
M. Al-Azawi and N. K. Ibrahim, "Bimodal
Histogram Based Image Segmentation Using
Fuzzy-Logic," in NCAIAE12, Massana, 2012.
4
H. Yang and X. Zhou, "Research of Content
Based Image Retrieval Technology," in
Proceedings of the Third International
Symposium on Electronic Commerce and
Security Workshops(ISECS ’10), Guangzhou,
P. R. China, 2010.
5
Y. Liu, D. Zhang, G. Lu and W.-Y. Ma,
"Asurvey of content-based image retrieval
with high-level semantics," Pattern
Recognition, no. 40, p. 262–282, 2007.
6
A. Koschan and M. Abidi, Digital Color
Image Processing, New Jersey: JOHN
WILEY & SONS, INC., 2008.
7
W. Skarbek and A. Koschan, Colour Image
Segmentation - A Survey, Sekretariat FR 5-4:
Leiter der Fachbibliothek Informatik, 1994.
8
S. Thilagamani and N. Shanthi, "A Survey on
Image Segmentation Through Clustering,"
International Journal of Research and
Reviews in Information Sciences, vol. 1, no.
1, pp. 14-17, March 2011.
9
T. N. Pappas, "An Adaptive Clustering
Algorithm for Image Segmentation," IEEE
TRANSACTIONS ON SIGNAL
PROCESSING, vol. 10, no. 1, pp. 901 - 914,
1992.
10
K. Deshmukh and G. N. Shinde, "An adaptive
neuro-fuzzy system for color image
segmentation," J. Indian Inst. Sci., no. 8, p.
493–506, Sept.–Oct. 2006.
11
J. Shi and J. Malik, "Normalized Cuts and
Image Segmentation," IEEE
TRANSACTIONS ON PATTERN ANALYSIS
AND MACHINE INTELLIGENCE,, vol. 22,
no. 8, August 2000.
12
A. Konar, Artificial Intelligence and Soft
Computing: Behavioral and Cognitive
Modeling of the Human Brain, Boca Raton
London New York Washington, D.C.: CRC
Press, 1999.
13
A. Gelbukh and R. Monroy, in Advances in
Artificial Intelligence / Research on
Computing Science, México, 2005.
14
J. Laaksonen, M. Koskela, S. Laakso and E.
Oja, "Self-Organising Maps as a Relevance
Feedback Technique in Content-Based Image
Retrieval," Pattern Analysis & Applications,
vol. 4, no. 2-3, p. 140–152, 2001.
15
C.-F. Tsai, K. McGarry and J. Tait, "Image
Classification Using Hybrid Neural
Networks," in ACM, Toronto, 2003.
16
S. Čabarkapa, N. Kojić, . V. Radosavlje and
B. Reljin, "Adaptive Content-Based Image
Retrieval with Relevance Feedback," in
EUROCON, Serbia & Montenegro, Belgrade,
2005.
17
N. S. Kojić, S. K. Čabarkapa, G. J. Zajić and
B. D. Reljin, "Implementation of Neural
Network in CBIR Systems with Relevance
Feedback," Journal of Automatic Control,
University of Belgrad, vol. 16, pp. 41-45,
2006.
18
B. Thomee and M. Lew, "Relevance feedback
in content-based image retrieval: promising
directions," in Proceedings of the 13th annual
conference of the Advanced School for
Computing and Imaging, Heijen,
Netherlands, 2007.
19
S. Nematipour, . J. Shanbehzadeh and R. A.
Moghadam, "Relevance Feedback
Optimization in Content Based Image
Retrieval Via Enhanced Radial Basis
Function Network," in Proceeding of the
International Multiconference of Engineers
and Computer Scientists IMECS, Hong Kong,
2011.
20
H. Lee and S. Yoo, "Intelligent image
retrieval using neural network," IEICE Trans.
on Information and Systems, Vols. E84-D, no.
12, pp. 1810-1819, 2001.
21
Y. Zhu, X. Liu and W. Mio, "Content-Based
Image Categorization and Retrieval using
Neural Networks," in IEEE International
Conference on Multimedia and Expo, Beijing
, 2007.
22
V. Balamurugan and P. Anandhakumar,
"Neuro-Fuzzy Based Clustering Approach
For Content Based Image Retrieval Using
2D-Wavelet Transform," International
Journal of Recent Trends in Engineering, vol.
1, no. 1, pp. 418-424, 2009.
23
S. Sadek, A. Al-Hamadi, B. Michaelis and
Sayed, "Cubic-splines neural network- based
system for Image Retrieval," in 16th IEEE
International Conference on Image
Processing (ICIP), Cairo , 2009 .
24
S. Sadek, A. Al-Hamad, B. Michaelis and U.
Sayed, "Image Retrieval Using Cubic Splines
Neural Networks," International Journal of
Video & Image Processing and Network
Security IJVIPNS, vol. 9, no. 10, pp. 17-22,
2010.
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
764

25
P. Rao, E. Vamsidhar, G. V. P. Raju, R.
Satapati and K. Varma, "An approach for
CBIR system through multi layer neural
network," International Journal of
Engineering Science and Technology, vol. 2,
no. 4, pp. 559-563, 2010.
Biography
Mohammad A. N. Al-Azawi: Mohammad Al-Azawi is
working as head of computer science and management
information system Department at Oman College of
Management and Technology. He has over 12 years of
teaching experience at Baghdad University and
Mustansiriyah University (Iraq), Yarmouk University
(Jordan), and Gulf College and Oman College (Oman).
Al Azawi is a member of many national and international
organizations like; the international society of Iraqi Scientists
(USA) and Iraqi Engineers Union (Iraq). Also he is an author
and reviewer in many periodicals and journals like
International Journal of Interactive Mobile Technologies
(iJIM) and International Journal of Advanced Corporate
Learning (iJAC) (Austria).
His researches interests include; e-environment, image
processing, computer vision and understanding, computer
intelligence..
APPENDIXES
Figure 8 shows the implementation of
the algorithm, in which the neural
network has been trained and the image
in entered to the neural network and the
NN recognized the image and gave the
reults as a stop sign.
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
765

Figure 8. The implementation of the algorithm
International Journal of Digital Information and Wireless Communications (IJDIWC) 1(4): 753-766
The Society of Digital Information and Wireless Communications, 2012 (ISSN 2225-658X)
766