Emblematical image based pattern recognition paradigm using Multi-Layer Perceptron Neural Network

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 6, Ver. II (Nov – Dec. 2015), PP 59-66
www.iosrjournals.org
DOI: 10.9790/0661-17625966 www.iosrjournals.org 59 | Page
Emblematical image based pattern recognition paradigm using
Multi-Layer Perceptron Neural Network
Md. Abdullah-al-mamun1
, Aminul Islam2
1, 2
(Computer Science and Engineering, Rajshahi University of Engineering and Technology, Bangladesh)
Abstract: The abstract Likewise human brain machine can be signifying diverse pattern sculpt that is
proficiently identify an image based object like optical character, hand character image, fingerprint and
something like this. To present the model of image based pattern recognition perspective by a machine, different
stages are associated like image acquiring from the digitizing image sources, preprocessing image to remove
unwanted data by the normalizing and filtering, extract the feature to represent the data as lower dimension
space and at last return the decision using Multi-Layer Perceptron neural network that is feed feature vector
from got the feature extraction process of a given input image. Performance observation complexity is discussed
rest of the description of pattern recognition model. Our goal of this paper is to introduced symbolical image
based pattern recognition model using Multi-Layer Perceptron learning algorithm in the field of artificial
neural network (like as human-like-brain) with best possible way of utilizing available processes and learning
knowledge in a way that performance can be same as human.
Keywords: Pattern Recognition; Multi-Layer Perceptron; MLP; Artificial Neural Network (ANN);
Backpropagation.
I. Introduction
“A picture is worth a thousand words.” As human beings, we are able to tell a story from a picture
based on what we see and our background knowledge [1]. Is it possible to ascertain a semantic concept from an
image by a machine that will be performing same as human like brain? In a short, the answer will be “possible”.
As human brain is capable to easily identify image based object as their learning knowledge is growing from
childhood. So if we are propose an artificial neural model that is perform like human-like-brain then it is
possible to detect an image object. To able a computer program in semantic understanding by recognition image
based pattern, it is need to extract visual feature efficient and effective way and build a model that is capable
learning background knowledge like human brain.
Layered neural networks involve a set of input cells connected to a collection of output cells by means of
one or more layers of modifiable intermediate connections [3]. The neural network architecture is associating by
an input pattern, that representation situation of a pattern, which produced an output class by feed the network. In
the context of image processing, the different stages are: acquisition, concern with digitizing the data coming
from a sensor such as camera or scanner, localizing, which involves extracting the object from its background,
and representation [2]. In the final decision stage, consist of two phases learning and recognition that are
associated with each other‟s. In the learning stage artificial neural network is adapted by the training of known
object classes. Network adaptation rate is proportional to the number of training that is performing on the
network. To perform recognition phase, object feature vector is feed into the network and according to the
previous learning knowledge of the adapted network is return a output feature vector class for the given object
feature vector.
II. Model of pattern recognition
An overall organization of the pattern recognition representation can be illustrated with the figure is as
following,
Figure-1: Model of Pattern recognition.

Emblematical image based pattern recognition paradigm using Multi-Layer Perceptron…
1.1. Image Acquiring
In the acquiring image part input the image document from any hardware device that can be scanned
copy image, optical character or something like this. It is the initial stage for the pattern recognition model
because without input how to classify?
1.2. Image Pre-processing
Image preprocessing system preprocesses the image both geometrically and radiometrically by the
radiant energy characteristics in such a way that produced the image of the original scenes. By the preprocessing
system reduced the noise of the image, normalize the image and filtering the image so that image can‟t contains
the unwanted data.
1.3. Feature Extraction
Feature extraction measured data and builds derived features intended to be informative, non-
redundant, facilitating the subsequent learning and generalization steps, in some cases leading to better human
interpretations [13]. By the performing operation of feature extraction on the image, it extracts the feature of an
image to produce the relevant data vector in a lower dimensional space from the input image.
1.4. Classification
The final part of the classification that is associated with the two parts; one is training process and
recognition process. In the training phase, network is adapted from the several training set of data with correct
class for a given training set. After the network is trained, classification for new glyphs can be done by
extracting features from new glyphs and using the trained network to predict their class [9].
III. Acquiring Image
The opening stage in the workflow progression of any vision system is performing image acquisition in
the digital image processing because of without an image there is no query is arise to process it. Image
acquisition is the branch of Digital Image Processing can be defined as the action of retrieval of an image from
any possible source, generally hardware-based source. After image has been acquired fittingly, there are several
preprocessing technique can be applied on the image to execute for the further processing. If the image has not
been acquired satisfactorily then the intended tasks may not be achievable, even with the aid of some form of
image enhancement [6].
IV. Preprocessing Image
In imaging science, image processing is processing of images using mathematical operations by using
any form of signal processing for which the input is an image, such as a photograph or video frame; the output
of image processing may be either an image or a set of characteristics or parameters related to the image [10].
This image preprocessing system produces a correct image both geometrically and radiometrically by the
radiant energy characteristics that is closest as possible of the original scenes.
1.5. Normalization
In the field of image processing normalization refer to the changes the range of pixel intensity values.
In the digital image processing it also referred to the dynamic range expansion. Applications include
photographs with poor contrast due to glare, for example. Normalization is sometimes called contrast stretching
or histogram stretching [10].
Normalization transforms an n-dimensional grayscale image with
intensity values in the range (Min, Max), into a new image with
intensity values in the range (newMin, newMax).
The linear normalization of a grayscale digital image is performed according to the formula
…………………………………………………...… (1)
According to the histogram analysis, if we see intensity range of the image is [50,180] and that should
be converted to the desired range are [0,255]. Now subtracting 50 from each of pixel intensity, that making the
range [0,130]. Then each pixel intensity is multiplied by 255/130 to making the range [0,255].
1.6. Filtering
In the field of digital signal processing, filtering is the process to remove the unwanted data,
component or feature signal from the original source of data. Filtering is a class of signal processing, removing
some frequencies and not others in order to suppress interfering signals and reduce background noise,

Sharpening and Deblurring. Correlations can be removed for certain frequency components and not for
others without having to act in the frequency domain [11].
V. Feature Extraction
In the task of pattern recognition it should be differenced the possible output classes according to the
input set of data. So the feature extraction is essentially done after preprocessing of classification phase in
pattern recognition and machine learning. In the process of pattern recognition input data is transformed into a
reduced set of representation of feature to produced correct output class.
In machine learning, pattern recognition and in image processing, feature extraction starts from an
initial set of measured data and builds derived values (features) intended to be informative, non-redundant,
facilitating the subsequent learning and generalization steps, in some cases leading to better human
interpretations [13]. The main goal of feature extraction is to extract set of feature to produce the relevant
information from the original input set data that can be represent in a lower dimensionality space.
1.7. Approach of Feature Extraction
Generally, an image contains some feature is as - color, texture, shape, edge, shadows, temporal and so
on. All features can be classified into two types level feature is as low-level features and high-level features.
Low-level features can be extracted directed from the original images, whereas high-level feature extraction
must be based on low-level features [8]. Here we should be most concern about color, texture and shape feature
that feature extraction details are following.
1.8. Color of Feature Extraction
The entire color feature of an image is broadly used to retrieval image feature space for the feature
extraction purpose. In general, the color of an image can be specified a color model in terms of 3-D coordinate
system. The color model can be RGB (red, green, blue), HSV (hue, saturation, value) and Y,Cb,Cr (luminance
and chrominance). The relationship can be defined is as following,
Figure-2: RGB and HSV color space.
1.9. Texture of Feature Extraction
Texture is a powerful regional descriptor of an image to retrieval feature extraction process. It is the
most important characteristic which has been used to classify and recognize objects and have been used in
finding similarities between images in multimedia databases [16]. Consider, the matrix element p(i,j) represent
the relative frequency of two neighbor pixel at distance d with grey levels „i‟ and „j‟ at a given direction. So,
Statistical methods based on the grey level co-occurrence matrix can be express,
……………………………………………………………………………
…………….…………….. (6)
Where, Ng represents the total number of gray level.

1.10. Shape of Feature Extraction
Shape of an image is a primitive feature for description of image content. Shapes are often the
archetypes of objects belonging to the same pattern class [4], and it can be as a silhouette (i.e, rotation, scale and
translation) of an object. Features calculated from objects contour: circularity, aspect ratio, discontinuity angle
irregularity, length irregularity, complexity, right-angleness, sharpness, directedness [16]. Those factors are
invariant to the shape descriptors.
VI. Multi-layer Perceptron
Artificial Neural Networks (ANNs) are a family of statistical learning models inspired by biological
neural networks (the central nervous systems of animals, in particular the brain) that interconnected to send
messages each other and also used to estimate or approximate functions that can depend on a large number of
inputs and are generally unknown [7]. In the branch of Artificial Neural Network, A Multi-layer
Perceptron(MLP) is the feedforward artificial neural network model that is design in a way that input set are feed
from the input layer and that will be produced an output vector set. A MLP consists of two or more layers where
each layer contains several number of neuron with nonlinear activation function and each neuron of one layer is
fully connected to the next layer as a direct graph (except input layer).
Figure-3: Sample architecture of Multi-Layer Perceptron Neural Network.
Consider, in the MLP has number of hidden layer is one then it can be express as the function f:RD
=>
RM
, where D is the input vector size and M is the output vector size y=f(x). Here, the output layer matrix
notation can be express as,
f(x)=G(bj+Wj(S(bi+Wix))
Where bi, bj are the bias vector for each neuron; Wi,Wj are the weight of the connected link and G,S are
the activation function.
The vector of the hidden layer can be express as so that weighted matrix
is connected the input vector to hidden layer. Here, sigmoid function is used to activation function that
can be express as
1.11. Learning
Learning is the process of a system that it enables the system to do that‟s types of task efficiently for
the next time. It is essential for unknown environments, i.e., when lack of output control to make decision. The
problem of learning is to take in prior knowledge and data (e.g., about the experiences of the agent) and to create
an internal representation (the knowledge base) that is used by the agent as it acts [18]. Appling the learning
process algorithm within the multilayer network architecture, the synaptic weights and threshold are update in a
way that the classification/recognition task can be perform efficiently. A class of functions training process
means fined using a set of observations which solves the task in some optimal sense [19]. Here to
perform the learning process for the multilayer perceptron neural network we have used error backpropagation
learning rule.
Consider a multilayer neuron network that has d dimensional input feature with M numbers of neuron
in the hidden layer which is produced k numbers of output vector. And sigmoid function is used as activation
function is
Now calculated the activation for Hidden layer bj is following,
Wji is the weight for the network connection from ith
level input layer to jth
level hidden layer where xi
is the feature input. Now transferring each activation by the non-linear activation function,

In the context of neural network, Zj interpreted the output of the hidden layer unit.
Now for the hidden layer, the Zj is linearly combined to the activation of the Kth
output layer.
Where, k=1,2,…,K
Transferring each activation by non-linear activation function is following,
These equations may be combined to give the overall equation that describes the forward propagation
through the network, and describes how an output vector is computed from an input vector, given the weight
matrices:
Now to get the output vector from the input vector, we have combined the equation is as following,
……………………………………………………………….…..(7)
That can be describing as a forward propagation throughout the network.
Figure-4: 3-Layer MLP with bias connection.
1.12. Learning through Backpropagation
Back-propagation learning rule perform of two sweeps called forward sweep and backward sweep
throughout different layer in the network. In forward sweep or forward pass, an activity pattern (input vector) is
applied to the sensory nodes of the network, and its effect propagates through the network layer by layer [17]. In
the time of forward pass, the synaptic (network link) weight are fixed and after pass the values throughout the
layer an output is produced as the actual response of the network. If the produced output is not equal to the
desired output then an error signal is produced and to reduce the error, it needs to perform backward pass
throughout the network. During the backward pass, the synaptic weights are adjusted in accordance with an
error-correction rule [17]. By the performing sweeps, the produced output error is propagated backward through
the network.
Now we can evaluate the training network using gradient descent function. If error function can be
define as E then derivation form is and
1.12.1. Error Function
To train an MLP we need to define an error function. Again we use the sum-of-squares error function,
obtained by summing over a training set of N examples:
For the training set N, the error function E can be defined as sum-of-square error function that is
summing over training sample N.
……………………………………………………………………… (8)
Where, ynk is the output vector that is get from the forward propagation and tnk is the targeted output for
the output layer.
So the gradient descent for the output layer is sum of overall training set,
……………………………………………………………………………….…….. (9)
Similarly, Gradient descent for the hidden layer,
……………………………………………………………………………..……….. (10)

1.12.2. Hidden-to-Output Weight
The error gradients for the hidden-to-output weights from equation (9) is,
…………………………………………………………………… (11)
Where, the error signal can be defined as derivation form of the gradient error En with respect to
the activation ank
………………………………………………….(12)
And,
1.12.3. Input-to-Hidden Weight
The error gradients for the input-to-hidden weights can be express as, . The error signal for hidden
unit j
……………………………………………………….….(13)
Here, because of ank is the weighted sum of Wkj. And znj is the first derivation
of bnj. So from the equation (13) we get,
………………………………………………………….……….(14)
This equation (14) is described as back-propagation error formula. So the derivation of input-to-hidden
weight can be express as,
……………………………………………………………….……..(15)
Equation (15) will be applied recursively for the next level hidden layer unit.
VII. Multi-layer Perceptron Learning Algorithm
1. Initializes all weights and thresholds with minimum random value.
2. Present input Xp = x0,x1,x2,…,xn-1 and target output Tp = to,tl,t2,. . . ,tm-l where n is the number of input nodes
and m is the number of output nodes. For classification, Tp is set to zero except for one element set to 1 that
corresponds to the class that Xp is in [20].
3. Calculate the output for each layer
This output will be input for the next layer. Where, layer has contains n numbers of nodes and w is the
weight for the each link.
4. If the targeted output is not equal the desire outputs then it need to apply backward sweep to reduce the
error level by updating the weight and threshold value.
5. Now adjust weight for the output layer can be calculated by,
Adjust threshold for the output layer can be calculated by,
6. Adjust weight for the hidden layer can be calculated by,
Adjust threshold for the hidden layer can be calculated by,
The backward sweep repeated until error difference between actual and desire value reduced within the
minimum level.
VIII. Complexity of Multi-layer Perceptron Learning Algorithm
The performance of pattern recognition can be influence by the Multi-Layer perceptorn neural network
for the following reasons is describe below,

1.13. Parameters of MLP
1.13.1. Numbers of Input
Increasing the number of input to feed the input layer can be increased the complexity of network
topological structure and may be influencing the overall network (like number of hidden layer) structure. That
may be caused the increased the error rate. This can be clear by an example is that if the number of input is 100
then to get an optimal topology network it need to about 300 neurons for the hidden layer!
1.13.2. Numbers of Hidden Layer
Significant change can be shown if the numbers of Hidden layer are increased in the network. For
example, to solve the XOR problem by 2-Layer and 3-Layer it can be classify is as following,
Layer Figure Classification
2-Layer
3-Layer
Table-1: Comparison between Network and Classification.
Local minima can be considered to occur when two or more disjoint classes are categorized as the
same. These amounts to a poor internal representation within the hidden units, and so adding more units to this
layer will allow a better recoding of the inputs and lessen the occurrence of these minima [20].
1.13.3. Number of Epochs
Increasing the number of epochs of the network can provide the better performance but many times it
may be return the poor result for the recognition system. This partially can be attributed to the high value of
learning rate parameter as the network approaches its optimal limits and further weight updates result in
bypassing the optimal state [19]. This process is known as over-learning.
1.13.4. Learning rate parameter
The variation of the learning rate parameter can be influence the network performance. If the learning
rate parameter value is high then network neurons and link values are updating more quickly comparing if the
learning rate parameter value is less to reach the optimal state.
1.14. Feature Extraction
Feature extraction plays a big role for the total operational process for any pattern recognition system
in the field of machine learning system. Because of the feature extraction process build a source raw data into
convert in a way that it should be informative, non-redundant and successive for the learning or recognition
process. So it should be apply better feature extraction process produced set of data that should be obtain a
linear or non-linear transformation of the original set of data.
1.15. Data Collection
Input source data is important each stages for the pattern recognition because this data will be used for
the further processing like feature extraction, classification and recognition of a pattern. So the data collection
method should be design in a way that input data is garbage or rush free and it is analyzable for the complex
multivariate techniques. To perform inception operation on the data it needs to check the quality of the data,
calculating the summary statistics of the data an sketch the plots of the data to get the structure of the data [19].
IX. Conclusion
Presenting the pattern recognition sculpt to identify image based object, we are present the multi-layer
perception learning algorithm for learning and classification method for decision making scheme with pre-
processing technique after getting the input image. The significant feature of the Multi-Layer Perceptron is that
the learning rates are dynamically computed each epoch by an interpolation map and error function is
transformed into a lower dimensional error space and the reduced error function is employed to identify the
variable learning rates [9]. In the field of artificial neural network, it is really difficult to recognize an image
based object by the multi-layer perceptron learning algorithm due to real time complexity of the pattern whereas
pattern can be linear or non-linear. Multi-Layer Perceptrons can be return 100% accurate result only for the

linearly separable problems. Whereas for the non-linear separable pattern is difficult to recognition. Because the
learning process of a multilayer perceptron requires the optimization of an error function E(y,t) comparing the
predicted output y, and the observed target t [15]. So a lot of efforts are associated to improving higher
recognition accuracy.
References
[1] Dong ping, Tian, "A Review on Image Feature Extraction and Representation Techniques", International Journal of Multimedia and
Ubiquitous Engineering, Vol. 8, No. 4, July, 2013, pp. 385-396.
[2] Sarvda Chauhan, Shalini Dhingra, "Pattern Recognition System using MLP Neural Networks", International Journal of Engineering
Research and Development, e-ISSN: 2278-067X, Volume 4, Issue 9 (November 2012), PP. 43-46.
[3] “A Versatile Machine Vision System for Complex Industrial Parts” IEEE Trans. On Computers, Sep. 1977, No. 9, Vol. C-27.
[4] L.daF. Costa, J.R.M. Cesar, “Shape Analysis and Classification: Theory and Practice”, CRC Press, Florida, USA, 2001
[5] Eugene P, Angela B, "What Is Image Acquisition in Image Processing", wisegeek.com, 2003-2015 Conjecture Corporation.
[6] A D Marshall, R R Martin, "Computer Vision, Models and Inspection", World Scientific Series in Robotics and Intelligent Systems:
Volume 4, ISBN: 978-981-02-0772-4 (hardcover), March 1992.
[7] En Wikipedia, “Artificial neural network”, 17 July 2015.
[8] E. Saber, A.M. Tekalp, ”Integration of color, edge and texture features for automatic region-based image annotation and retrieval,”
Electronic Imaging, 7, pp. 684–700, 1998.
[9] Dharm Singh, Neha Mehta, Pallavi Purohit,"Text Based Image Recognition using Multilayer Perceptron", IP Multimedia
Communications, A Special Issue from IJCA - www.ijcaonline.org, pp. 143-146.
[10] Rafael C. Gonzalez; Richard E. Woods (2008). “Digital Image Processing”, Prentice Hall. pp. 1–3. ISBN 978-0-13-168728-8.
[11] en Wikipedia, “Filter_(signal_processing)”, 4 July 2015.
[12] Gaurav Kumar,Pradeep Kumar Bhatia, "A Detailed Review of Feature Extraction in Image Processing Systems",2014 Fourth
International Conference on Advanced Computing & Communication Technologies, pp. 5-12, Febuary 2014.
[13] En Wikipedia, “Feature_extraction”, 7 July 2015.
[14] Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze, “Introduction to Information Retrieval”, Cambridge University
Press. 2008, ISBN. 0521865719.
[15] Luís M. Silva, J. Marques de Sá, Luís A. Alexandre, "Data classification with multilayer perceptrons using a generalized error
function", ScienceDirect, elsevier.com, Neural Networks 21 (2008) 1302-1310, pp. 1302-1310.
[16] Ryszard S. Chora´s, "Image Feature Extraction Techniques and Their Applications for CBIR and Biometrics Systems",
International journal of biology and biomedical engineering, pp. 6-16 ,Issue 1, Vol. 1, 2007.
[17] Amrita Hirwani, Sandeep Gonnade, "Character Recognition Using Multilayer Perceptron", Amrita Hirwani et al, / (IJCSIT)
International Journal of Computer Science and Information Technologies, Vol. 5 (1) , 2014, pp. 558-661.
[18] David Poole, Alan Mackworth, "Artificial Intelligence", Cambridge University Press, 2010.
[19] Md. Abdullah-al-mamun, Mustak Ahmed, "Hypothetical Pattern Recognition design using Multi-Layer Perceptorn Neural Network
for supervised learning", Proc. in International Journal of Advanced Research, 2015.
[20] R Beale, T Jackson, "Neural Computing: An Introduction", Bristol, Philadelphia and New York.

Emblematical image based pattern recognition paradigm using Multi-Layer Perceptron Neural Network

More Related Content

What's hot (19)

Similar to Emblematical image based pattern recognition paradigm using Multi-Layer Perceptron Neural Network (20)

More from iosrjce (20)

Recently uploaded (20)

Emblematical image based pattern recognition paradigm using Multi-Layer Perceptron Neural Network