Vous êtes sur la page 1sur 4

International Journal of Image Processing And Visual Communication

Volume 1,Issue 1,August 2012



6

A Neural Network based Character Recognition System for
Indian Standard High Security Number Plates
Fajas F.
1
, Farhan Yousuf
1
, Remya P. R.
1
, Sajan Ambadiyil
2
, Varsha S
3


1. Cochin University of Science and Technology, Cochin
2. Optical Image Processing Team, Centre for Development of Imaging Technology, Thiruvananthapuram
3. Fourth Dimension, Chennai

*
anprindia@gmail.com


Abstract Automatic Number Plate Recognition or
ANPR is a mass surveillance method that uses optical character
recognition on images to read the number plates on vehicles. This
system is designed with a neural network which is trained to
recognize all the characters that can be found in an Indian
Standard High Security Number Plate and is implemented using
MATLAB.

KeywordsNumber Plate, Indian Standards, ANPR, OCR,
MATLAB
I. INTRODUCTION
The ANPR work is generally framed into the steps Number
plate extraction, character segmentation and character
recognition. From the entire input image, only the number
plate is detected and processed further in the next step of
character segmentation. In character segmentation phase each
and every character is isolated and segmented. Based on the
selection of prominent features of characters, each character is
recognized, in the character recognition phase.
Many countries have implemented ANPR systems to their
various surveillance systems because they have
standardization in number plates. Very soon, Indian number
plates on vehicles are also going to get standardized with the
introduction of High Security Number Plates for uniform
display of Indian registration. It is in this regard this system
Automatic Number Plate Recognition for Indian Standards is
built.

Fig.1 A Sample of Indian Number Plate
II. DESCRIPTION OF THE SYSTEM
The captured image taken from 5-12 feet away from
the vehicle mounted with Indian Standard High Security
Number Plate. It is then processed through the number plate
localiser and gives its output to segmentation part.
Segmentation part separates the characters individually. And
finally recognition part recognizes the characters giving the
result as the plate number. In simple terms this system takes
the image of a vehicle with Indian standard number plate as its
input and gives the characters displayed in the form of
computer data as its output.


Fig.2 System Design
III. PRE-PROCESSING
A. Plate Localisation
Localising (identifying a number plate) is an
algorithmic function that determines what aspect of the
vehicle's image is the number plate. This variance can further
compound the complexity for an algorithm to ascertain what
area of a vehicle constitutes a number plate and what area is
not. The algorithm searches for a similar background colour of
unified proportion and contrast as a means to differentiate
objects on a vehicle.

International Journal of Image Processing And Visual Communication
Volume 1,Issue 1,August 2012

7

The captured image is initially converted into gray
scale and edge detection is applied on it. Possible number
plate areas in the image are identified by observing sudden
changes in contrast. Remaining areas are filtered out. The best
possible number plate location is found out by comparing
width by height factor of actual Indian number plates to the
same factor of plate like areas found by this method. This
system showed maximum efficiency when the width by height
factor was set between 3 and 7.Fig 3.a, Fig 3.b, Fig 3.c, Fig
3.d shows the various stages of number plate localisation.


B. Character Segmentation
This Algorithm looks for characters of equal colour
and equidistance, with similar font structures to break apart
each individual character. Apart from the characters, noises
can also be segmented by this algorithm and treated as an
output. The characters are identified properly by the Optical
Character Recognition (OCR) phase with the help of a neural
network learning and recognition system.
IV. NEURAL NETWORK FOR OCR
Optical Character Recognition is an electronic translation
of scanned documents into machine readable/editable format.
For matching the characters with the database, input images
must be equal-sized with the database characters. Here the
characters are fit to size in database. The extracted characters
cut from plate and the characters on database are now equal-
sized. The next step is template matching. Template matching
is an effective algorithm for recognition of characters. The
character image is compared with the ones in the database and
the best similarity is measured. A neural network based
approach is used to implement this phase. A learning
mechanism is used to create the database and a recognition
mechanism for template matching.
A. Learning Mechanism
A simplified architecture of artificial neural networks is
used in this system. In the used method, characters are taught
to the neural network in a supervised manner. A character is
presented to the system and is assigned a particular label.
Several variant patterns of the same character are taught to the
neural network under the same label. Hence the neural
network learns various possible variations of a single
character and becomes adaptive in nature. During the training
process, the input to the neural network is the input matrix M
defined as follows:

If I (i, j) = 1 Then M(i, j) = 1
Else:
If I (i, j) = 0 Then M(i, j) = -1 (1.1)

In the current method of learning, each candidate
character taught to the network possesses a corresponding
weight matrix. For the character sample k to be taught to the
software network, the weight matrix is denoted by Wk. As
learning of the system progresses, it is this weight matrix that
is updated. At the commencement of teaching (supervised
training), this matrix is initialized to zero. Whenever a
character sample is to be taught to the neural network, an
input pattern representing that character is submitted to the
neural network. The neural network is then instructed to
identify this pattern as, say, the k in a knowledge base of
characters. That means that the pattern is assigned a label k. In
accordance with this, the weight matrix Wk is updated.

for all i = 1 to x
{
for all j = 1 to y
{
Wk(i,j) = Wk(i,j) + M(i,j)
}
} (1.2)
Here x and y are the dimensions of the matrix Wk (and M).
We can notice certain peculiarities in such a weight matrix,
like,
1. The matrix-elements with higher (positive) values are the
ones which stand for the most commonly occurring image-
pixels.
2. The elements with lesser or negative values stand for pixels
which appear less frequently in the images.
The weights may represent the importance or priority of a
parameter, which in the instant case is the occurrence of a
particular pixel in a character pattern. It can be seen that the
weights of the most frequent pixels are higher and usually
positive and those of the uncommon ones are lower and often
negative. The matrix therefore assigns importance to pixels on
the basis of their frequency of occurrence in the pattern. In
other words, highly probable pixels are assigned higher
priority while the less-frequent ones are penalized. However,
International Journal of Image Processing And Visual Communication
Volume 1,Issue 1,August 2012

8

all labelled patterns are treated without bias, so as to include
impartial adaptation in the system.
B. Recognition Mechanism
A network is constructed with each node trained to
recognize a particular character .The weight matrix is updated
for each of node. When a character is produced for
recognition, the input is fed to every node and tested for match
.Decision is taken in favor of a stored character depending on
the Candidate Score, Ideal Weight-Model Score and
Recognition Quotient (Q).
i. Candidate Score: This statistic is a product of corresponding
elements of the weight matrix Wk of the kth learnt pattern and
an input pattern I as its candidate. It is formulated using the
formula as follows:

(1.3)

Unlike in the training process where M was the processed
input matrix, in the recognition process, the binary image
matrix I is directly fed to the system for recognition.
ii. Ideal Weight-Model Score ():
This statistic simply gives the sum total of all the positive
elements of the weight matrix of a learnt pattern. It may be
formulated as follows

For i=1 to x
{
For j = 1 to y
{
If Wk(i,J) > 0 then
{
(k) = (k) + Wk(i,j)
} (1.4)

iii. Recognition Quotient (Q): This statistic gives a measure of
how well the recognition system identifies an input pattern as
a matching candidate for one of its many learnt patterns. It is
simply given by:
(1.5)

The greater the value of Q, the more confidence does the
system bestow on the input pattern as being 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 recognition
quotient (Q (k)) for each learnt pattern k.
2. Determine the value of k for which Q (k) has the maximum
value.
3. Too low maximum value of Q (k) (a threshold obtained
after multiple trials) indicates poor recognition. In such a case:
Conclude that the candidate pattern does not exist within the
knowledge base
OR
Teach the candidate pattern to the network till a satisfactory
value of Q (k) is obtained
4. Conditionally, identify the input candidate pattern as being
akin to the kth learnt pattern or proceed with the training for
better performance.
V. RESULT AND DISCUSSIONS
The proposed system had been simulated using MATLAB.
The number plates shown on Fig 4.a, Fig4.b, and Fig 4.c were
used for the learning mechanism and create the database for
the characters.


VI. CONCLUSION
A simple and efficient system has been developed to
localize the number plates from the image of a captured
vehicle containing Indian standard number plate .A neural
network based character recognition system has been
implemented to identify all the characters that can be found in
an Indian standard number plate. The system is found to have
good performance in comparing and matching the test pattern
with already stored patterns. System is adaptive and gives
satisfactory results in case of slight variation in the same
characters due to noise. The system threshold was found to be
0.85 which can be further improved on training. For same
character the recognition quotient is as high as 0.937 while for
different character it is lower than 0.5 hence the system
provides well distinction in case of different characters.
REFERENCES
[1] Avinash Kumar Jha, Supriya Narasimham, Sudheer Sreedhara Krishna,
V.P. Mahadevan Pillai, A Neural Network Based approach for
Fingerprint recognition system, 2010
International Journal of Image Processing And Visual Communication
Volume 1,Issue 1,August 2012

9

[2] Prathamesh Kulkarni, Ashish Khatri, Prateek Banga, Kushal Shah,
Automatic Number Plate Recognition (ANPR) System for Indian
conditions
[3] Alexander J. Faaborg, Using Neural Networks to Create an Adaptive
Character Recognition System, March 2002
[4] K.M. Sajjad, Automatic License Plate Recognition using Python and
OpenCV

Vous aimerez peut-être aussi