The automatic license plate recognition(alpr)

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 02 Issue: 08 | Aug-2013, Available @ http://www.ijret.org 353
THE AUTOMATIC LICENSE PLATE RECOGNITION(ALPR)
Rinku Solanki1
, Rajeshkumar Rai2
, Teena Raikwar3
1
M.tech student, digital communication, NRI-IST, Bhopal, Madhya Pradesh, India,
2
Assistant professor, Professor, Head of the department of E.C, NIIST, Bhopal, Madhya Pradesh, India
3 Assistant professor, electronica & communication dept., (NRI-IST, Bhopal), Madhya Pradesh, India
rinkusolanki86@gmail.com, raj.rai1008@gmail.com, t_raikwar.2006@yahoo.co.in
Abstract
Every country uses their own way of designing and allocating number plates to their country vehicles. This license number plate is
then used by various government offices for their respective regular administrative task like- traffic police tracking the people who are
violating the traffic rules, to identify the theft cars, in toll collection and parking allocation management etc. In India all motorized
vehicle are assigned unique numbers. These numbers are assigned to the vehicles by district-level Regional Transport Office (RTO).
In India the license plates must be kept in both front and back of the vehicle. These plates in general are easily readable by human due
to their high level of intelligence on the contrary; it becomes an extremely difficult task for the computers to do the same. Many
attributes like illumination, blur, background color, foreground color etc. will pose a problem.
Index Terms: Automatic license plate recognition (ALPR) system, proposed methodology, reference
-----------------------------------------------------------------------***-----------------------------------------------------------------------
1. INTRODUCTION
The purpose of this paper is to provide researchers a systematic
survey of existing ALPR research by categorizing existing
methods according to the features they used, by analyzing the
pros/cons of these features, and by comparing them in terms of
recognition performance and processing speed, and to open
some issues for the future research.
Basic block diagram of the ALPR system is shown in fig 1.for
above steps different techniques used by different author which
are studied in literature review. An example of the number
plate extraction is given .by this figure block diagram is easily
understand, in this figure all steps of block diagram is shown
by indicating number A, B, C, D.
Automatic license plate recognition (ALPR) applies image
processing and character recognition technology to identify
vehicles by automatically reading their number plates.and this
system mainly devide in three steps:all ateps are better explain
in proposed methodology.
FIGURE (1) BASIC BLOCKDIAGRAM OF ALPR
SYSTEM
The variations of the plate types or environments cause
challenges in the detection and recognition of license plates.
They are summarized as follows:
1) Location: Plates exist in different locations of an image.
2) Quantity: An image may contain no or many plates.
3) Size: Plates may have different sizes due to the camera
distance and the zoom factor.
4) Color: Plates may have various characters and
background colors due to different plate types or capturing
devices.
5) Font: Plates of different nations may be written in
different fonts and language.
6) Occlusion: Plates may be obscured by dirt.
7) Inclination: Plates may be tilted.
8) Other: In addition to characters, a plate may contain
frames and screws.
Environment variations:
1) Illumination: Input images may have different types of
illumination, mainly due to environmental lighting and
vehicle headlights.
2) Background: The image background may contain
patterns similar to plates, such as numbers stamped on a
vehicle, bumper with vertical patterns, and textured floors.
2. PROPOSEDMETHODOLOGY
Vehicle license plate (VLP) constitutes an unambiguous
identifier of a vehicle participating in road traffic. Reading a
license plate is the first step in determining the identities of
parties involved in traffic incidents. An efficient automatic
license plate recognition process may become the core of fully

__________________________________________________________________________________________
computerized road traffic monitoring systems, electronic fee
collection solutions, surveillance devices and safety
supervision systems. It is important that the recognition
accuracy of such a process is very high. Tracking and
registering dangerous behavior in traffic may be used for
prosecuting offenders.
License-Plate Recognition System consists of three
main modules:
(1) License plate detection
(2) Character segmentation
(3)Character Recognition (CR)
Step(1)image acquisition, in image acquisition explained that
from where images are acquire Image can be input to the
system by different methods by analog camera,or by digital
cameras, but nowadays digital technology has their advantages
so better input method is by digital cameras or by direct digital
photos.
2.1 License Plate Detection
by whole capturing image we having license plate covered by
background of vehicle body,so by this step only plate are is
extracted from whole body. our task now is to identify the
region containing the license plate. In this experiment, two
features are defined and extracted in order to decide if a
candidate region contains a license plate or not , these features
are
2.1.1 Preprocessing
Since we have assumed that the license plate has a yellow
background, the first step is to identify the regions in the image
that contain the intensity of RGB corresponding to the color
yellow. We Filter the yellow colored part from the image using
values obtained by experiments on the 10 sample images. (i) -
(a< R< b) && (p< G< q) && (x< B< y) Where R is the
intensity of the color Red , G of Green and B that of
Blue.Based on this condition we obtain a Binary Image. We
change yellow to white and non-yellow to black.
2.1.2 Morphological Operations:
These are Non-linear filters, with the function of restraining
noises, extracting features and segmenting images etc The
following morphological operations have been used
Fill (MATLAB function – imfill): fills holes in the binary
image. A hole is a set of background pixels that cannot be
reached by filling in the background from the edge of the
image.
Open (MATLAB function – imopen): performs morphological
opening on the grayscale or binary image with the pre-defined
structuring element.
Dilate (MATLAB function – imdilate): dilates the grayscale,
binary, or packed binary image returning the dilated image
(I)On applying these morphological operations on „im1.jpg‟
we obtain the Image.
2.1.3 Horizontal Segmentation
Once the Preprocessing is over, the next step is to segment the
license plate candidates from the image. We first do a
horizontal segmentation of the image using the histogram
method.
Steps:
For this we calculate the horizontal and vertical projections of
intensity. Then we find the mean of the local minimas of
horizontal projection plot .Based on the threshold calculated
from the above local minimas, we find x locations of the
segmented regions. In order to locate the right and left edges of
License plate from candidate region, the vertical projection
after mathematical morphology deal is changed into binary
image. The arithmetic for doing this is:
TfLf TTT ≥= 1)( TfLf TTT <= 0)(
Where fT (1,i) is the vertical projection after mathematical
morphology, T is the threshold. Then scan the function of fT
(1,i) and register the portions where values change from 0 to 1
and from 1 to 0 in stack1 and stack2 respectively. So the
candidate position of the left and right edge of the license plate
are in stack1 (1,i) and stack2(1,i) respectively, and the
candidate’s width of the license plate is calculated by: width(1,
i) stack2(1, i) - stack1(1, i) respectively. So the candidate
position of the left and right edge of the license plate are in
stack1 (1,i) and stack2(1,i) respectively, and the candidate’s
width of the license plate is calculated by: width(1, i) stack2(1,
i) - stack1(1, i) These give the x coordinates of the potentially
candidates regions
2.1.4 Getting Potential Candidates
After getting the horizontal segments of the candidate regions,
we would now want to get the vertical coordinates, in order to
extract the exact area from the image For each of the horizontal
segments, in order find the vertical location, we once again use
the vertical projections of intensity. Then converting to a
binary image using the threshold as discussed in the previous
section we get the desired vertical coordinates. Now we have
all our candidate regions. An candidate regions obtained from
our sample image „im1.jpg‟ are :

__________________________________________________________________________________________
Figure (2) Candidate regions
2.1.5 Identifying the License Plate:
Out of the many candidate regions that we have obtained from
the so far mentioned procedure, our task now is to identify the
region containing the license plate. In this experiment, two
features are defined and extracted in order to decide if a
candidate region contains a license plate or not , these features
are
1. Aspect ratio
2. Edge Density
Even though these features are not scale-invariant, luminance-
invariant, rotation-invariant, but they are insensitive to many
environment changes.
Aspect ratio
The aspect ratio is defined as the ratio of the width to the height
of the region.
Aspect Ratio = width/height
Since the minimum enclosing rectangle(MER) of the object
region can be computed via rotating the region in previous
section, the dimension of the object‟s MER can be taken as the
width and the height of the region.
Edge density
Applying the above feature to filter the segmented regions, a
lot of nonlicense plate regions can be removed. However, there
are still many candidate regions left which take similar
rectangularity and aspect ratio features as the license plate
regions do, such as often the head lights. Considering that the
license plate regions generally take higher local variance in its
pixels‟ values due to the presence of characters, an important
feature to describe license plate region is local variance, which
is quantized using the edge density. The edge density is
measured in a region R by averaging the intensities of all edge
pixels within the region as
∑∈
=
Rnm
D
nmE
NR
R ,
1 ),(
Where E(i,j) represents the edge magnitude at location (i,j), and
NR is the number of pixels in region R. License plate Sets used
for training to calculate aspect ratio and edge density: We used
24 license plates as training data. They are contained in the
directory named
“edge_density_and_aspect_ratio_training_images”.
After performing this experiment, we found the average values
of the above mentioned features as follow:
Aspect ratio = 4.4912 edge density = 0.1359
On using the above mentioned features we are able to drop the
incorrect regions and get the final result as the extracted license
plate from the input image.
For the candidate images shown above, above tests rules out
the first candidate and gives the following output –
Figure (3) Number Plate
2.2 SEGMENTATION
This step characters on license plate are segmented and
identify. This step is the most important step in license plate
recognition because all further steps rely on it. This is the
second major part of the License Plate detection algorithm.
There are many factors that cause the character segmentation
task difficult, such as image noise, plate frame, rivet, space
mark, plate rotation and illumination variance. We here
propose the algorithm that is quite robust and gives
significantly good results on images having the above
mentioned problems for the segmentation preprocessing is
required by conversion to gray scale and binarization. Different
algorithms are used for segmentation which are explained
further later in literature review.
In the segmentation of plate characters, license plate is
segmented into its constituent parts obtaining the characters
individually. Firstly, image is filtered for enhancing the image
and removing the noises and unwanted spots. Then dilation
operation is applied to the image for separating the characters
from each other if the characters are close to each other. After
this operation, horizontal and vertical smearing are applied for
finding the character regions. The result of this segmentation is
in Figure given below.
Figure (4) Locations of plate characters
The next step is to cut the plate characters. It is done by finding
starting and end points of characters in horizontal direction.
The individual characters cut from the plate are as follows in
Figure given below.

__________________________________________________________________________________________
Figure (5) Number plate characters
2.3 CHARACTER RECOGNITISON:
Automatically locate license plates by principal visual word
(PVW), discovery and local feature matching. Observing that
characters in different license plates are duplicates of each
other, we bring in the idea of using the bag-of words (Bow)
model popularly applied in partial-duplicate image search.
Unlike the classic Bow model, for each plate character, we
automatically discover the PVW characterized with geometric
context. Given a new image, the license plates are extracted by
matching local features with PVW. Besides license plate
detection, our approach can also be extended to the detection of
logos and trademarks. Due to the invariance virtue of scale-
invariant feature transform feature, our method can adaptively
deal with various changes in the license plates, such as rotation,
scaling, illumination, etc. Promising results of the proposed
approach are demonstrated with an experimental study in
license plate detection. We formulate license plate detection as
a visual matching problem. For each character, we collect SIFT
features falling into the character region and generate PVW by
unsupervised clustering. The amount of PVW for each plate
character is determined automatically Besides SIFT
descriptors, each visual word contains some geometric
information, such as orientation, ratio of scale to character
height, and relative position in the character region. Those
geometric clues will be used to filter false feature matches and
estimate the character and plate size. In testing, every valid
match votes a support for plate location, and all supports are
unified to discover potential license plates. Due to the
invariance virtue of SIFT feature, our method can adaptively
deal with various changes of license plate, such as distortion
from observation views, scaling, and illumination. Multiple
license plates in a single image can also be automatically
detected.
In visual word matching and license plate locating, we compare
the extracted SIFT features of the test image with all
discovered PVW, and locate the license plate based on the
matching results. Let us denote the PVW set as {D, G} ={(di ,
gi ), i = 1, . . . , N}, where di denotes appearance descriptor,
and gi denotes the geometric clues, N denotes the visual word
number. Once the PVW of an object category is discovered, we
can use it for detection in a new image. Given features F = { fi
} for a test image, the probability that the test image
corresponds to a sign of interest is
)0,(
),(
)0()0,(
),0( ii
ii
ii
ii gdp
gdp
pgdp
gdp α
⋅
=
……………………………. (1)
Where p(O) is prior of plate. The likelihood p(di , gi |O) is
deduced as
∑ ⋅=
j
jjiiii fpfgdpgdp )0()0,,()0,(
……………………………. (2)
Where p(di , gi | f j , O) is modeled by matching feature f j to
the descriptor di of the PVW.
Figure (6) Illustration of a PVW (red arrow) in the character
“6”
Consequently, by searching for the local maxima of the
likelihood function given by (1) for all PVW, we can find the
initial hypotheses for license plate location. Some other prior
heuristics can also be imposed to remove potential false
positives. In the following subsections, we will discuss PVW
generation and local feature matching to extract license plate in
detail.
A. PVW Generation
There are a certain number of sorted characters in license
plates, each with the same format, but maybe undergoing
illumination change or affine transformation. Since SIFT
feature is invariant to changes in scale and rotation, and robust
to illumination change and affine distortion [2], some
repeatable and distinctive SIFT features to each character exist,
called PVW. As shown in Fig., a PVW is denoted as V (des,
ori, rat, pos), where des is the 128-D SIFT descriptor, ori is the
SIFT orientation (−π ≤ ori < π), rat = H/s (s is the SIFT scale),
and pos = ( f/W, e/H) is a 2-D vector denoting the relative
position of the key point in the character region. Both des and
ori are originated from the standard SIFT features [2] des
captures the local visual appearance with a concatenation of 8-
D orientation histograms from 4 by 4 subpatches around local
interest point. ori denotes the dominant directions of local
gradients around a key point. Relative to ori and des is
represented to achieve invariance to image rotation changes .
Ideally, for a feature with high repeatability in a certain
character, rat shall be identical. Given this specific SIFT
feature with scale s, we can estimate the corresponding
character height as rat · s. Given an image patch of the
character with height, v, we can also derive the scale of the
SIFT feature as v/rat. We collect many training images, each

__________________________________________________________________________________________
containing one or more license plates. License plates in the
training images are all upright, with little affine distortion.
Each character in the license plate is annotated and all SIFT
features in each character region are extracted. Usually, many
noisy features also exist. To discover the PVW of each
character, we need to cluster the local features of each
character and discover the most representative cluster centers
as the PVW, which can be found automatically in the
clustering process. In affinity propagation, a similarity matrix
of samples shall be defined. We first give the distance metric,
which will be used to define the similarity metric. The distance
between two feature samples V j and Vk is defined in (3)
PrOdkj DDDDd ⋅+⋅+⋅+⋅= δγβα,
…..……..(3)
where α, β, γ, and δ are constant weighting factors, Dd ,Do, Dr,
and Dp are the distance of descriptor, orientation, height-scale
ratio, and position, respectively, and are defined as follows:
∑=
−=
128
1
2
)(
1
i
k
i
j
id desdesD
σ
……………… (4)
( )kjkj
O orioriorioriD −−−⋅= π
π
2,min
1
……………. (5)
kj
r ratrat
N
D −=
1
…………. (6)
∑=
−=
2
1
2
)(
2
1
i
k
i
j
ip posposD
………………... (7)
Where σ and N are normalization factors to make sure that both
Dd and Dr range from 0 to 1.
The similarity metric is a decreasing function of the distance
metric. There are many choices for it. In our implementation,
the pair wise similarity between two feature samples V j and
Vk is defined as
)0(,)( ,, >−= ndS n
kjkj
(8)
In affinity propagation, the diagonal elements in the similarity
matrix are referred to as exemplar preference, which will
influence the number of identified clusters. Generally, without
any priori, we set it as the median of the input similarities, after
clustering; we need to discover the most representative clusters.
For each cluster, we count the number of image patches which
contain at least one feature falling into the cluster. Then an
image-number histogram is built. To select those representative
clusters, a threshold thresh shall be specified on the histogram.
Any cluster with image number above thresh will be selected.
In each selected cluster, the PVW are defined as the average of
all samples falling into that cluster. In our experiments, we set
thresh = 0.6·Num, where Num is the total sample number of
the specific character. Fig. illustrates the feature clustering
results of three characters: “0,” “6,” and “9.” In each character,
the PVW are highlighted in red color on the patch with its
geometric information: ori, rat, and pos. The PVW of
characters from “0” to “9” are shown in Fig. while those of
characters from “A” to “Z” excluding “I” and“O” are
illustrated in Fig. 5. There are no PVW of character “I” and
“O” as these two characters are not found in any training plate.
B. Visual Word Matching
Given a test image, we will discover those characters with
features matched to the PVW. We first extract SIFT features
from the test image. Then each SIFT feature F(des, ori, scl) is
compared with the PVW of each character. A feature is
considered as a candidate match if the minimum descriptor
distance to a certain PVW of a certain character is less than
constant threshold T
{ } Tdesdestd d
t
F
td
≤−=
2,
**
argmin,
(9)
Where desF denotes the descriptor vector of a test SIFT
feature, des d t denotes the descriptor vector of the t-th PVW in
the d-th plate character. In standard SIFT features, descriptors
are all normalized to be a constant const. In our experiments,
we set T = 0.5 · const. Each candidate match is recorded as
C(x, y, angle, height, pos), where x and y denote spatial
position of the test SIFT feature in the image plane, angle is the
rotation angle from the test feature to the matched visual word,
height = rat dt* · sclF is the estimated height of the
corresponding license plate, pos = pos d*t* denotes the relative
position in character.
Figure (7) PVW of each digit in Chinese license plate each
arrow denotes a PVW

__________________________________________________________________________________________
Figure (8) PVW of each letter in Chinese license plate, from
“A” to “Z,”excluding “I” and “O” each arrow denotes a PVW.
2.3. License Plate Locating
Once the character features in the test image are identified, we
can make use of the geometric context of the matched PVW to
locate the license plate. A bounding box will be estimated to
encompass license plate by determining the upper, lower, left,
and right bounding lines sequentially.
We first estimate the upper and lower bounding lines of license
plate in images. Specially, each matched feature C(x, y, angle,
height, pos) estimates a point (xup, yup) of the upper bounding
line by (10) and (11)
)sin(cos)2( θθ −⋅⋅−= hposxxuo (10)
)cos(sin)2( θθ +⋅⋅−= hposyyup
(11)
Where h and θ are the median of height and angle of all valid
matched features C, respectively, pos(2) = e/H. The origin is
assumed to be at the upper-left corner of the image plane.
Then, for all upper bounding points, we estimate a line with
linear regression. Similarly, we can also determine the lower
bounding points (xdown, ydown) with (12) and (13) and
estimate the lower bounding line of license plate.
)sin(cos))2(1( θθ −⋅⋅−+= hposxxdown (12)
)cos(sin))2(1( θθ +⋅⋅−+= hposyydown (13)
After that, we can also roughly estimate the left and right
bounding lines. In license plate, the ratio of plate width to
height is constant. When the plate height h is estimated, we can
obtain the plate width w. Since license plate must cover all
matched key points of SIFT feature, the interval between the
left bound and the most right key point of matched feature shall
be no less than w, so is that of the interval between the right
bounding line and the most left key point. Consequently, we
can determine the minimal bounding box containing the license
plate although some background patch is also included, they
can be removed with some other information, such as edge
map. However, this is not our focus in this paper.
3. EXPERIMENTAL RESULTS
To evaluate the proposed approach containing 410 Chinese
license plate images. Of them, 160 license plate images are
downloaded from the Internet while another 250 images are
taken by the authors. The second dataset contains 112 images
with resolution of 896 ?592, each contains a U.S. license plate
with a cluttered background, such as trees or grass. In this
dataset, the plate character height ranges from about 16 to 23.
On the first dataset, it achieves a 93.2% “true” detection rate,
On the second dataset, the “true” detection rate of our approach
is 84.8%,All the “true” detection rates of four approaches are
lower than that on the first dataset. This is because the plates in
the second dataset are of smaller size, and the background is
much more cluttered. The false positive rate of our approach is
also much lower than that of the three comparison approaches.
We investigate the time efficiency from two aspects. The first
one is feature extraction time, and the second one is detection
time after feature extraction. Our approach is based on SIFT
feature, whose extraction time cost is larger than that of edge
maps, as used in the other three comparison approaches. The
detection time cost of our approach is proportional to the SIFT
feature amount of image, while the detection efficiency of the
three comparison approaches are determined by the complexity
extent of image texture. In open environment, there are various
observation views from cameras, which will make the edge
map-based methods difficult to accurately extract the whole
plate. However, benefiting from the invariance property of
SIFT feature, our approach can effectively address that
difficulty when the observation angle is within some tolerance
range. Some results seen in figure12
Figure (9) Sample detection results of license plates
CONCLUSIONS
In general, an ALPR system consists of four processing stages.
In the image acquisition stage, some points have to be
considered when choosing the ALPR system camera, such as
the camera resolution and the shutter speed. In the license plate
extraction stage, the license plate is extracted based on some
features such as the color, the boundary, or the existence of the
characters. In the license plate segmentation stage, the
characters are extracted by projecting their color information,
by labeling them, or by matching their positions with template.
Finally, the characters are recognized in the character
recognition stage by template matching, or by classifiers such

__________________________________________________________________________________________
as neural networks and fuzzy classifiers. Automatic license
plate recognition is quite challenging due to the different
license plate formats and the varying environmental conditions.
There are numerous ALPR techniques have been proposed in
recent years.
REFERENCES:
[1] Muhammad Sarfraz, Mohammed Jameel Ahmed, and Syed
A. Ghazi ,“Saudi arebian licence plate recognition system”,
International Conference on Geometric Modeling and Graphics
(GMAG’03), 2003.
[2] Serkan Ozbay, and Ergun Ercelebi,” Automatic Vehicle
Identification by Plate Recognition”, Processing of world
academy of science engineering and technology vol9, ISSN
1307-6884, november 2005.
[3] Humayun Karim Sulehria, Ye Zhang, Danish Irfan, Atif
Karim Sulehria,” Vehicle Number Plate Recognition Using
Mathematical Morphology and Neural Networks”, WSEAS
TRANSACTIONS on COMPUTERS, Volume 7,ISSN: 1109-
2750, Issue 6, June 2008.
[4] Zhen-Xue Chen, Cheng-Yun Liu, Fa-Liang Chang, and
Guo-You Wang,” Automatic License-Plate Location and
Recognition Based on Feature Salience”, IEEE Transaction on
vehicle technology, VOL. 58, NO. 7, september 2009.
[5] Dr. P.K.Suri, Dr. Ekta Walia, Er. Amit Verma,” Vehicle
Number Plate Detection using Sobel Edge Detection
Technique”, International Journal of Computer Science and
Technology, ISSN : 2229 – 4333, IJCST Vol. 1, Issue 2,
December 2010.
[6] Kumar Parasuraman, Member, IEEE and P.Vasantha
Kumar, “ An Efficient Method for Indian Vehicle License Plate
Extraction and Character Segmentation”, IEEE International
Conference on Computational Intelligence and Computing
Research,2010.
[7] Muhammad H Dashtban, Zahra Dashtban, Hassan Bevrani,
“ A Novel Approach for Vehicle License Plate Localization
and Recognition”, International Journal of Computer
Applications (0975 – 8887), Volume 26– No.11, July 2011.
[8] Stuti Asthana, Niresh Sharma, Rajdeep Singh, “Vehicle
number plate recognition using multiple layer back propagation
neural networks”, International Journal of Computer
Technology and Electronics Engineering (IJCTEE), Volume 1,
Issue 1, July 10, 2011.
BIOGRAPHIES:
Rinku K. Solanki, M.Tech student (digital communication),
NRI-IST, Bhopal, Madhya Pradesh, India,
Rinkusolanki86@gmail.com
RAJESH KUMAR RAI, Asst. Prof.(Head of the department
of E.C.), NIIST, Bhopal, Madhya Pradesh, India
raj.rai1008@gmail.com
Teena Raikwar, Assistant professor, electronica &
communication dept., (NRI-IST, Bhopal), Madhya Pradesh,
India
t_raikwar.2006@yahoo.co.in

The automatic license plate recognition(alpr)

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to The automatic license plate recognition(alpr) (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

The automatic license plate recognition(alpr)