Visual Character Recognition using

Artificial Neural Networks

the domain of
ABSTRACT artificial
intelligence. In this
paper, an
The computing world introduction to
has a lot to gain from nueral networks
neural networks. Their and a simplified
ability to learn by neural approach to
example makes them recognition of
very flexible and optical or visual
powerful They are also characters is
very well suited for real portrayed and
time systems because of discussed.
their fast response and
computational times
which are due to their
parallel architecture.
The recognition of
optical characters
is known to be one
of the earliest
applications of
Artificial Neural
Networks, which
partially emulate
human thinking in
 knowledge
1. through learning.
INTRODUCTION: 2. A neural
A neural network is an network's
interconnected group of knowledge is
neurons. A biological stored within
neuron may have as inter-neuron
many as 10,000 different connection
inputs, and may send its strengths known
output to many other as synaptic
neurons. Neurons are weights.
wired up in a 3-
dimensional pattern.
The prime examples are
PURPOSE:
of biologically neural
networks, especially the A neural network is a
human brain. These may mathematical or
be considered as self computational model for
learning mechanisms information processing
which don't require the based on a connectionist
traditional skills of a approach to
programmer. computation. The
In modern usage, original inspiration for
researchers often refer to the technique was from
an artificial neural examination of
network (ANN) simply bioelectrical networks in
as a “neural network” or the brain formed by
“neural net” for short. neurons and their
These ANN resemble synapses. In a neural
the human brain in the network model, simple
following way: units (or processing
elements ) are connected
1. A neural network
together to form a
acquires
network of nodes - ARTIFICIAL
hence the term "neural NEURON
network". These
processing elements are
termed as “neurons” or
“nodes”.

Neural networks can

simulate some
intelligence activities of
the human brain, such as
sense and inspiration,
It is observed that the
as well as thinking of
evolving of ANN causes
images.
it to eventually reach a
The Artificial Neuron is state where all neurons
actually quite simple. continue working but no
All signals can be 1 or further changes in their
-1 (the "binary" case, state happen. A network
often called classic spin may have more than one
for its similarity with the stable states, and it is
problem of disordered obviously determined by
magnetic systems). The the choice of synaptic
neuron calculates a weights and thresholds
weighted sum of inputs for the neurons.
and compares it to a
T
threshold. If the sum is
he recognition of
higher than the
threshold, the output is characters from

set to 1, otherwise to -1. scanned images of

documents has
been a problem
that has received
much attention in
the fields of image deduced from
processing, pattern samples.
recognition and As such, the
artificial human system of
intelligence. vision is excellent
Classical methods in the sense of the
in pattern following qualities:
recognition do not 1. The human brain
as such suffice for is adaptive to
the recognition of minor changes and
visual characters errors in visual
due to the following patterns. Thus we
reasons: are able to read the
1. The ‘same’ handwritings of
characters differ in many people
sizes, shapes and despite different
styles from person styles of writing.
to person and even 2. The human
from time to time vision system
with the same learns from
person. experience: Hence
2. Like any image, we are able to
visual characters grasp newer styles
are subject to and scripts with
spoilage due to amazingly high
noise. speed.
3. There are no 3. The human
hard-and-fast rules vision system is
that define the immune to most
appearance of a variations of size,
visual character. aspect ratio, color,
Hence rules need location and
to be heuristically
orientation of visual compared with
characters. greater challenges
In such as recognition
contrast to of human faces.
limitations of ANNs have enjoyed
classical considerable
computing, success in this area
Artificial Neural due to their
Networks (ANNs), humanoid qualities
that were first such as adapting to
developed in the changes and
mid 1900’s serve learning from prior
for the emulation of experience. The
human thinking in subsequent parts of
computation to a the paper elucidate
meager, yet this fact in more
appreciable extent. details.
Of the several
fields wherein they The
have been applied, paper is organized
humanoid as follows: in
computing in section [2.], image
general and digitization, which
pattern recognition is an essential step
in particular have prior to neural
been of increasing networking, is
activity. The described. Section
recognition of [3.] describes the
visual (optical) learning
characters is a mechanism of the
problem of neural network
relatively amenable used, and the
complexity when employed
architecture is alone. For example,
described in it may contain
section [4.]. pictures and colors
Section [5.] that do not provide
discusses the any useful
issues that affect information in the
the performance of instant sense of
the proposed character
methods with recognition. In
reference to its addition,
accuracy, characters which
computational need to be singly
complexity and analyzed may exist
extensibility. as word clusters or
may be located at
[2.] IMAGE various points in
DIGITIZATION: the document.
Such an image is
When a usually processed
document is put to for noise-reduction
visual recognition, and separation of
it is expected to be individual
consisting of characters from the
printed (or document. It is
handwritten) convenient for
characters comprehension to
pertaining to one or assume that the
more scripts or submitted image is
fonts. This freed from noise
document however, and that individual
may contain characters have
information besides already been
optical characters located (using for
example, a suitable 6X8=48 digital
clustering cells, each having a
algorithm). This single color, either
situation is black or white. It
synonymous to the becomes important
one in which a for us to encode
single noise-free this information in
character has been a form meaningful
submitted to the to a computer. For
system for this, we assign a
recognition. value +1 to each
black pixel and 0 to
each white pixel
and create the
binary image
matrix I which is
shown in the Fig.
(1.c). So much of
The process conversion is
of digitization is enough for neural
important for the networking which is
neural network described next.
used in the system. Digitization of an
In this process, the image into a binary
input image is matrix of specified
sampled into a dimensions makes
binary window the input image
which forms the invariant of its
input to the actual dimensions.
recognition system. Hence an image of
In the above figure, whatever size gets
the alphabet A has transformed into a
been digitized into binary matrix of
fixed pre- network in a
determined supervised manner.
dimensions. This A character is
establishes presented to the
uniformity in the system and is
dimensions of the assigned a
input and stored particular label.
patterns as they Several variant
move through the patterns of the
recognition system. same character are
taught to the
network under the
[3.] LEARNING same label. Hence
MECHANISM: the
network learns
In the various possible
employed system, variations of a
a highly simplified single pattern and
architecture of becomes adaptive
artificial neural in nature. During
networks is used. the training
For purpose of easy process, the input
understanding, the to the neural
learning network is the
mechanism of the input matrix M
neural network is defined as follows:
described first and If (,) 1 I i j = Then
its architecture is (,)1Mij=
described next, in Else:
section [4.]. In the If ( , ) 0 I i j = Then
used method,
( , ) 1 M i j = − (1.1)
various characters
The input matrix M
are taught to the
is now fed as input
to the neural matrix is initialized
network. It is to zero. Whenever
typical for any a character is to be
neural network to taught to the
learn in a network, an input
supervised or pattern
unsupervised representing that
manner by character is
adjusting its submitted to the
weights. In the network. The
current method of network is then
learning, each instructed to
candidate identify this pattern
character taught to as, say, the kth
the network character in a
possesses a knowledge base of
corresponding characters. That
weight matrix. For means that the
the kth character to pattern is assigned
be taught to the a label k. In
network, the accordance with
weight matrix is this, the weight
denoted by Wk. As matrix Wk is
learning of the updated in the
character following manner:
progresses, it is
this weight matrix for all i=1 to x
that is updated. At {
the
for all j=1 to y
commencement of
{
teaching
(supervised Wk(i,j)=Wk(i,j
training), this )+M(I,j)
 } Fig.

} (3) gives the

weight matrix, say,
WS corresponding
to the alphabet S.
Here x and y are
The matrix is has
the dimensions of
been updated
the matrix Wk (and
thrice to learn the
M). The following
alphabet S. It
figure shows the
should be noted
digitization of three
that this matrix is
input patterns
specific to the
representing S that
alphabet S alone.
are presented to
Other characters
the system for it to
shall each have a
learn.
corresponding
weight matrix.

Note that the

patterns slightly
differ from each
other, just as
handwriting differs
from person to
person (or time to
time) and like A close observation
printed characters of the matrix would
differ from machine bring the following
to machine. points to notice:
1. The matrix- in the instant case
elements with is the occurrence of
higher (positive) a particular pixel in
values are the ones a character
which stand for the pattern. It can be
most commonly seen that the
occurring image- weights of the most
pixels. frequent pixels are
2. The elements higher and usually
with lesser or positive and those
negative values of the uncommon
stand for pixels ones are lower and
which appear less often negative. The
frequently in the matrix therefore
images. assigns importance
to pixels on the
Neural basis of their
networks learn frequency of
through such occurrence in the
updating of their pattern. In other
weights. Each time, words, highly
the probable pixels are
weights are assigned higher
adjusted in such a priority while the
manner as to give less-frequent ones
an output closer to are penalized.
the desired output However, all
than before. The labeled patterns
weights may are treated without
represent the bias, so as to
importance or include impartial
priority of a adaptation in the
parameter, which system.
 done on the basis
of certain statistics
[4.] NETWORK that shall be
ARCHITECTURE: defined next.

The overall
architecture of the
recognition system
is shown in Fig. (4). (4.1) Candidate
In this system, the Score ( ψ ):
candidate pattern I Thi
is the input. The s statistic is a
block ‘M’ provides product of
the input matrix M corresponding
to the weight elements of the
blocks Wk for each weight matrix Wk of
k. There are totally the kth learnt
n weightblocks for pattern and an
the totally n input pattern I as
characters to be its candidate. It is
taught (or already formulated as
taught) to the follows:
system.

It should be noted
that unlike in the
training process
where M was the
processed input
matrix, in the

The recognition of recognition

patterns is now process, the binary

image matrix I is
directly fed to the well the recognition
system for system identifies
recognition. an input pattern as
a matching
(4.2) Ideal candidate for one
Weight-Model of its many learnt
Score ( µ ): patterns. It is
simply given by:
This
statistic simply
gives the sum total
of all the positive
elements of the
The greater the
weight matrix of a
value of Q, the
learnt pattern. It
more confidence
may be formulated
does the system
as follows (with ( )
bestow on the input
k µ initialized to 0
pattern as being
each time).
similar to a pattern
already known to it.
The classification of
input patterns now
follows the
following trivial
procedure:-
1. For an input
candidate pattern I,
calculate the
(4.3) Recognition
recognition
Quotient (Q):
quotient ( ( ) Q k )
This
for each learnt
statistic gives a
pattern k.
measure of how
2. Determine the making the best
value of k for which selection as in Step
( ) Q k has the 4 of the
maximum value. aforementioned
3. Too low algorithm. The
maximum value of adaptive
( ) Q k (say less performance of the
than 0.5) indicates network can easily
poor recognition. In be tested by an
such a case: example: we
• Conclude that the submit two hand-
candidate pattern drawn patterns
does not exist representing S and
within the P respectively to
knowledge base OR the system that
• Teach the has already learnt
candidate pattern only the character
to the network till a S. The recognition
satisfactory value quotient yielded by
of ( ) Q k is the trained system
obtained. is
4. Conditionally, mentioned
identify the input alongside.
candidate pattern
as being akin to the
kth learnt pattern
OR proceed with
the training for
better
performance.
In Fig. (4),
the selector gives
an output k by
Note that the
pattern in Fig. (5)
does not exactly
appear like the
three patterns of
Fig. (2) that were
taught to the
system. However, The system
being adaptive, the however dismisses
system the candidature of
nevertheless the pattern
bestows a good representing P in
quotient 0.68 Q Fig. (5) by yielding
= on the pattern, a low value of
indicating a match. ( 0.21) Q = . It can
To improve be observed by
recognition of this regular teaching,
particular pattern, that the system
the same pattern develops on its
can be repeatedly ability to identify a
input to the system matching pattern
and taught to it as and reject
before under the nonmatching
same label. As a patterns. Thus,
result, the value of regular supervised
Q approaches unity teaching marks
after each time the enhanced
pattern is taught. performance of the
This illustrates system.
learning from prior
experience in [5.]
neural networks. PERFORMANCE
ISSUES:
 them as per the
The neural orientation of the
system has some user.
direct advantages
that become
apparent at this The dimensions of

stage: the input matrix

1. The method is need to be

highly adaptive; adjusted for

recognition is performance.

tolerant to minor Greater the

errors and changes dimensions, higher

in patterns. the resolution and

2. The knowledge better the

base of the system recognition. This

can be modified by however increases

teaching it newer the time-

characters or complexity of the

teaching different system which can

variants of earlier be a sensitive issue

characters. with slower

3. The system is computers.

highly general and Typically, 32X32

is invariant to size matrices have been

and aspect ratio. empirically found

4. The system can sufficient for the

be made recognition of

userspecific: User- English handwritten

profiles of characters. For

characters can be intricate scripts,

maintained, and greater resolution

the system can be of the matrices is

made to recognize required.

 described. The
As advantages of
already illustrated neural computing
in the previous over classical
example, efficient methods have been
supervised outlined. Despite
teaching is the computational
essential for the complexity
proper involved, artificial
performance. neural networks
Neural expert offer several
systems are advantages in
therefore typically pattern recognition
used where human- and classification in
centered training is the sense of
preferred against emulating adaptive
rigid and inflexible human intelligence
system-rules. to a small extent.

[6.]
CONCLUSION:

A simplistic
approach for
recognition of
visual characters
using artificial
neural networks
has been