Journal of Transportation Technologies, 2012, 2, 13-21 Published Online January 2012 (
A Design Flow for Robust License Plate Localization and
Recognition in Complex Scenes
Dhawal Wazalwar, Erdal Oruklu, Jafar Saniie
Department of Electrical and Computer Engineering, Illinois Institute of Technology, Chicago, USA
Received September 22, 2011; revised October 24, 2011; accepted November 8, 2011
In this paper, we present a new design flow for robu st license plate localizatio n and recognition. The alg orithm consists
of three stages: 1) license plate localization; 2) character segmentation; and 3) feature extraction and character recogni-
tion. The algorithm uses Mexican hat operator for edge detection and Euler number of a binary image for identifying
the license plate region. A pre-processing step using median filter and contrast enhancement is employed to improv e the
character segmentation performance in case of low resolution and blur images. A unique feature vector comprised of
region properties, projection data and reflection symmetry coefficient has been proposed. Back propagation artificial
neural network classifier has been used to train and test the neural network based on the extracted feature. A thorough
testing of algorithm is performed on a database with varying test cases in terms of illumination and differen t plate con-
ditions. Practical considerations like existence of another text block in an image, presence of dirt or shadow on or near
license plate region, license plate with rows of characters and sensitivity to license plate dimensions have been ad-
dressed. The results are encouraging with success rate of 98.10% for license plate localization and 97.05% f or char acter
Keywords: License Plate Localization; Character Recognition; Reflection Symmetry Coefficient; Artificial Neural
1. Introduction
License plate recognition is considered to be one of the
fastest growing technologies in the field of surveillance
and control. Apart from its increasing use in areas like
road traffic management and speed checks, it is also
becoming an important tool for police officials in solving
traditional criminal investigations. With such an exten-
sive use of this technology, efforts are constantly being
made to make it more efficient. The technology research
can be classified into two categories, namely fixed cam-
era and mobile camera license plate recognition. Each of
these categorias its own specific requirements and tech-
nical challenges. Our main focus in this paper is on fixed
camera applications.
License plate localization is very crucial step in that
the overall system accuracy depends on how accurately
we can detect the exact license he plate location. The
input can be in the form of still images or video frames
from surveillance cameras. The processing can be done
at either color or grayscale level. One such method based
on using unique color combination of license plate region
was proposed in [1]. However, the large data that needs
to be processed in case of color image processing makes
it unsuitable for real time applications. On the other
hand, the amount of data that needs to be handled in case
of gray scale technique is comparatively less and thus
facilitates faster and more efficient implementations. A
method based on using texture information of license
plate region as a feature [2] was proposed with an accu-
racy of 98.5%. However failures for cases with existence
of another text block, dirty license plate region in the
input frame were documented [2]. In order to make the
texture information of license plate clearer, a technique
based on obtaining horizontal image difference was pre-
sented in [3]. This method however is sensitive to the
license plate dimensions and is not robust enough to
handle all practical conditions. A real time solution with
intelligent frame selection from input video was pre-
sented in [4]. It then enhances the obtained frame using
Gaussian filter and histogram equalization. Morphologi-
cal operations are then done along with connected com-
ponent analysis to extract the license plate region. Mor-
phological operations, specifically dilation and erosion,
were used in [5] to detect license plate region. This tech-
nique in case of complex images may require an initial
system to get the part of image having vehicle separated
from the rest of the image. This can be particularly diffi
cult in cases where complete vehicle is not visible or
where license plate exists at corner parts in an image.
opyright © 2012 SciRes. JTTs
The main objective in character segmentatio n is to ob-
tain clean and isolated characters. Vertical projection in-
formation can be used to separate the characters [6].
However, this technique requires prior knowledge of
license plate size as well as the number of characters in it.
This sensitivity to license plate dimensions was reduced
to an extent by employing region growing segmentation
method along with pixel connectivity information [7]. In
order to tackle tilted and distorted license plates, image
tilt correction and gray level enhancement was proposed
[8]. Binarization of a gray level image is a key aspect in
character segmentation and researchers have been using
both local as well as global thresholding [7] as a means
of achieving it. Hybrid binarization technique using his-
togram information [9] improved the overall performance
especially in cases with presence of dirt or improper
shadows on the license plate.
The accuracy of a character recognition algorith m dep-
ends on uniqueness of feature vector information. Hori-
zontal and vertical projection data, zoning density and
contour features were combined to form a feature vector
[10] with an accuracy of 95.7%. Recently, a method
based on the concept of stroke direction and elastic mesh
[8] was proposed with an accuracy in the range of 97.8
99.5%. Another important aspect of character recognition
step is the type of classifier employed. A two stage clas-
sifier based on Euclidean distance calculation is defined
in [10]. Apart from various types of probability models,
classifiers based on artificial neural network (ANN) [11]
and support vector machine (SVM) [4,8] are wid ely used
in the field of optical character recognition.
In this paper, we present a new design flow for a ro-
bust license plate localization and recognition system,
shown in Figure 1. The system can be categorized into
three key stages 1) license plate localization; 2) character
segmentation; and 3) character recognition. In Section 2,
the main objective was to come up with an optimal li-
cense plate localization algorithm which will address to
some of the shortcomings in previous implementations
like 1) presence of another text block in an image [2]; 2)
existence of partial contrast between the license plate
region and surrounding area; 3) license plate region lo-
cated anywhere in the frame. In Section 3, the emphasis
is on obtaining clean and isolated characters. Practical
issues like presence of shadow or dirt on license plate,
blur images have been highlighted. In Section 4, main
focus is on obtaining unique feature vector as well as on
training the back propagation artificial neural network
classifier. Finally, in order to have common benchmark-
ing process [12,13] and to determine accuracy of license
Figure 1. Design flow for robust license plate localization and recognition system.
Copyright © 2012 SciRes. JTTs
plate detection step, we have tested our algorithm on
different types of practical images. A detailed summary
of observed results has been presented in Section 5. The
images are taken from a commercial City Sync’s auto-
matic number plate recognition camera video [14].
2. License Plate Localization
A detailed flow diagram highlighting the key steps in-
volved in the license plate recognition system is shown in
Figure 1. The main steps involved in license plate local-
ization are edge detection, morphological dilation opera-
tion and region growing segmentation.
2.1. Selection of Edge Detection Technique
In order to avoid any possibility of false license plate
detection, it is important to obtain clean and continuous
edges of license plate region. Ideally, these edges should
not be connected to any other surrounding parts. General
gradient operators lik e Prewitt, So b el operators are us eful
in cases when there is a clear and distinct contrast be-
tween the license plate and region surrounding it, as seen
in Figure 2.
However, in cases when there is a partial contrast,
these general gradient operators fail to give the desired
output. For example, in Figure 3(b) it can be seen that
the license plate region is connected to the surrounding
region. In order to avoid such cases, more advanced op-
erators like Mexican hat operator have to be used, see
Figure 3(c). Mexican hat operator first performs smoo-
thing operation and then extracts edges. This function is
also called as Laplacian of Gaussian (LoG) and mathe-
matically can be expressed [15] as
22 r
where r2 = x2 + y2 and σ is the standard deviation.
Masks of different sizes 3 × 3, 5 × 5 and 9 × 9 were
tested and the best results were seen for 9 × 9 operator.
Following is the 9 × 9 operator used in our implementa-
000 111000
1336136 33
1336136 33
000 111000
 
 
 
(a) (b)
Figure 2. (a) Sample image 1; (b) Edge detection using Pre-
witt operator with license plate edges clearly separated from
surrounding region.
(a) (b)
Figure 3. (a) Sample image 2; (b) Failure case of Prewitt
operator with license plate edges connected to surrounding
region; (c) Edge detection output using Mexican hat opera-
tor with clear and distinct edges.
2.2. Morphological Dilation Operation
If the license plate image is blurred, edge detection out-
put can have discontinuity as seen in Figure 4(b). This
can result in failure cases or incorrect detection output,
(see Figure 4(c)). In order to prevent such cases and
make these edges thick and co ntinuous similar to Figure
4(d), a dilation operation is performed. The improved
detection performance can be seen Figure 4(e).
Dilation operation can be mathematically expressed
[15] as
where X is the object and B is the structuring element.
The size and nature of structuring element is important
since smaller size can negate the desired effect and larger
Copyright © 2012 SciRes. JTTs
(a) (b)
(c) (d)
Figure 4. (a) Blur image (b) Edge detection output (c) Plate
detection output without dilation step (d) Dilation output (e)
Modified plate detection output with dilation step.
size can cause the license plate regions to be connected
to the surrounding area. For our algorithm, best results
were achieved with 3 × 3 ones matrix as a structuring
2.3. Region Growing Segmentation
Region grow ing is a procedure in which we group pix els
based on some predefined pixel connectivity information
to form sub regions. The performance of region growing
algorithm depends on the starting points and the stopping
rule selected for the implementation [15]. In our case,
region growing process was performed on whole image
so as to segment the entire image into differen t sub parts.
For every bright pixel, its neighboring four pixel con-
nectivity information was checked and region was grown
based on this information. In order to increase the al-
gorithm speed, sub-sampling was used, where algorithm
operation was performed for every two pixels, instead of
performing for each and every pixel in the image. The
final output will be several sub regions labeled based on
pixel connectivity information, see Figure 5.
2.4. Detecting the License Plate Region
After segmenting the entire image in smaller regions, fol-
lowing criteria can be used to identify license plate re-
gion from it.
Criterion 1: License plat e dimensions
Generally license plate dimensions are fixed for a par-
ticular state/country and so it can be used to determine
the license plate region. However, by doing this we make
algorithm sensitive to license plate dimensions, which
depend on the distance between the vehicle an d camera.
Criterion 2: Euler number of a binary image
Characters used in license plate region are alphanu
(a) (b)
Figure 5. (a) Output after dilation step; (b) Region growing
segmentation output.
meric (A-Z and “0-9”). If we carefully observe these
characters in an image, this can be seen as closed curves
like in 0, 9, 8, P etc. Thus clearly license plate region will
have more closed curves than in any other part of an im-
age. Euler number of an image can be used to distinguish
the different regions obtained after the region growing
output [16]. Euler number of a binary image gives the
number of objects minus the closed curves (holes). Ex-
perimentally, it was found that in alphanumeric regions,
value of Euler number is zero or negative, while for other
regions it has a positive valu e. Using Euler number crite-
rion will overcome the sensitivity issue encountered in
using Criterion 1. However, if the vehicle image has al-
phanumeric characters in more than one region algorithm
may fail; see Figure 6(a).
Criterion 3: Combi nati on of Crit erion 1 and Criteri on 2
In this implementation, we have used a combined ap-
proach of Euler number criterion as well as license plate
dimensions to overcome problem raised in case of Crite-
rion 2. Instead of predefining exact license plate dimen-
sions, we specify a range of acceptable minimum and
maximum license plate sizes. This ad dresses the sensitiv-
ity issue raised in Criterion 1. The modified result can be
seen in Figure 6(b).
3. Character Segmentation
After locating the license plate, the next important step is
to segment each character individually, avoiding any
possibility of joint ch aracters. Figure 1 shows all the ne-
cessary steps that need to be performed in order to ensure
accurate character segmentation.
3.1. Image Preprocessing
3.1.1. Median Filtering
Median filter is a non-linear spatial filter and is used
widely for image denoising [17,18]. In median filtering,
we replace the original pixel data by median of pixels
contained in a prescribed window. In order to ensure that
the edges of the characters are preserved, analysis was
done to determine the optimal median filter size. In this
case, best results were obtained for 3 × 3 filter window.
Copyright © 2012 SciRes. JTTs
(a) (b)
Figure 6. (a) Failure case for Criterion 2; (b) Criterion 3
output for part (a) case.
3.1.2. Contrast Enhancement
After observing the pixel da ta in the license plate region,
it was found that the character information for most of
the images in database is usually in the intensity range of
0 - 0.4 on a scale of 1. In order to obtain further contrast
enhancement, we increased the dynamic range for pixels
in this (0 - 0.4) intensity range.
Figure 7 shows application of all the above steps on
one such sample case. In Figure 7(b), we can see the en-
hanced version of original license plate image.
3.2. Threshold Operation
For further processing, we need to convert grayscale im-
age into a binary image by using a threshold operation.
There are basically two types of threshold operation:
3.2.1. Global Thresholding
In this case, we select one single threshold value for en-
tire image. Mathematically it can be represented as
1,( ,)
(,)0, otherwise
xy T
where f(x,y) is the original grayscale image, g(x,y) is the
threshold output image and T is the threshold value
This operation is easy to implement and also less time
consuming. However, it is sensitive to illumination con-
ditions and fails in cases if there is a brighter or darker
region other than object. One such case is shown in
Figure 8(b), where we can see character “O” and “2” are
connected because of unwanted dark region in original
grayscale image, see Figure 8(a).
3.2.2. Local Thresholding
In the local threshold operation, we divide the grayscale
image into several parts by using a windowing operation
and then perform the threshold operation separately for
each case. Since, we perform threshold operation for
each region separately; a single brighter or darker region
cannot corrupt the threshold value. Local thresholding
(a) (b)
Figure 7. (a) Input lice nse plate image be fore preprocessing;
(b) Enhanced image after preprocessing.
(a) (b)
Figure 8. (a) License plate image with unwanted dark re-
gion; (b) Global thresholding output, with joint characters;
(c) Local thresholding output with separated characters.
ensures that in most cases, we get clear and separated
characters. In Figure 8(c), we can see all the characters
are separated compared to Figure 8(b), where global
thresholding was performed. The performance of local
threshold operation will vary depending on the selected
window size. Experimentally, best results were obtained
when the window size was chosen approximately equal
to the general individual character size in the license
plate image. For our algorithm, the threshold value for
local region is obtained using Otsu’s method for gray
level histogram [19].
3.3. Morphological Erosion Operation
In certain situations if some dirt or shadow is present on
the license plate region, the output of local threshold op-
eration may not be sufficient. For example in Figure 9(b)
it can be seen even after the local threshold operation so-
me amount of unwanted noise is still present near char-
acter “4” and “X”. This noise can be removed by per-
forming morphological erosion operation (see Figure
9(c)). Mathematically, erosion operation can be repre-
sented [15] as
BxBXΘ (4)
where X is the object and B is the structuring element.
The size and nature of structuring element should be se-
lected with an aim of keeping the character information
intact and only removing unwanted parts in the binary
3.4. Region Growing Segmentation
The pixel connectivity information can be used to separate
Copyright © 2012 SciRes. JTTs
(a) (b)
(c) (d)
Figure 9. (a) License plate with shadow on characters 4 and
X; (b) Output after local threshold operation; (c) Output
after morphological erosion operation; (d) Separated char-
acters after region growing segmentation.
the characters. Region growing segmentation along with
four pixel connectivity approach is used in this case.
Figure 9(d) shows the separated characters after perfor-
ming region growing segmentation.
4. Character Recognition
After segmenting the characters in license plate, the next
important step is to extract feature information from the
character and recognize it.
4.1. Feature Extraction
The robustness of character recognition algorithm de-
pends on how well it can recognize different versions of
a single character. In Figure 10, all the images are for
same character G of size 24 × 10, however it varies in
terms of spatial information, noise content and thickness
of edges. The feature vector should be selected in such a
way that algorithm can still accurately recogn ize all these
varied representations of a single character.
In addition, sometimes it is difficult to distinguish be-
tween characters like D and 0 or 5 and S because of the
similarity in their appearance as well as spatial infor-
mation. In order to deal with these challenges, our algo-
rithm uses a combination of spatial information as well
region properties. Following is the description of some of
these key features.
Feature 1: This feature comprises of region properties
like area, perimeter, minor-major axis length, orientation
and Euler number of binary image. Orientation is the
angle b etween th e x-ax is and the major axis o f the ellipse
that has same secondary moments as that of region. An
analysis of variation of all region properties with respect
to the characters was performed and these six properties
were selected from them.
Feature 2: The second part of feature vector includes
projection information about X and Y axis. This forms
major part of feature vector and is unique for each char-
Figure 10. Different versions of character “G”.
Feature 3: Some of the characters like 0, M, N are
symmetric about X-Y axes. This information can be used
to distinguish different characters especially 0 and D.
This symmetry information can be quantified by a coef
ficient called as reflection symmetry coefficient. For 2D
images, a way of calculating this coefficient is proposed
in [20]. In this implementation for a character of size 24
× 10, we calculate measure of symmetry about X and Y
axis in following way
1) Consider Cin as the original character matrix where
information is stored in binary format. The center of ma-
trix is considered as the origin. Cx and Cy are the trans-
formed matrix obtained after flipping about the X and Y
axes respectively.
2) Calculate the area of original matrix Cin.
3) Determine the union of transformed matrixes Cx and
Cy with original matrix Cin.
4) Calculate the areas of regions (CinCx) and (CinCy).
5) The reflection coefficient can be determined as
in x
in y
Feature 3 has significantly improved algorithm’s per-
formance in dealing with complex cases like D and 0.
Experimentally it was found that for 0, the value of X
coefficient and Y coefficient are approximately similar.
On the other hand, since D has symmetry only along X
axis, value of X coefficient is larger compared to Y coef-
4.2. Artificial Neural Network Design
Artificial neural networks (ANN) are considered a po-
werful tool for solving complex engineering problems
like pattern classification, clustering/categorization, predic-
tion/forecasting etc. They are in general classified into
two categories namely feed-forward networks and recur-
rent (or feedback) networks. In this implementation, we
are using a feed-forwar d based neural network desi gn .
A simplest form of feed-forward neuron model is call-
ed as “Perceptron”. In this model, (x1, x2, ···, xn) are the
input vectors, (w1, w2, ···, wn) are the weights associated
with the input vectors, h is the summation of all inputs
with its weights and y is the output. The selection of ac-
Copyright © 2012 SciRes. JTTs
tivation function is decided by the complexity of pro-
posed problem. For our design, we have used a log sig-
moid activation function. The basic algorithm for this
model can be explained [21] as follows.
1) Initialize the associated weights and the activation
function for the model.
2) Evaluate the output response for given input pattern
,,, t
3) Update the weights based on following rule
  
wtwtd yx 
In case of back-propagation model, the error at the
output is propagated backwards and accordingly the
weights for inputs are adjusted. This concept can be ex-
tended to multiple layer model based upon the requi-
Following is the description of some of these key de-
sign parameters in reference to character recognition
4.2.1. Input Layers
Input layers to the neural network are the feature vectors
extracted for the all the characters. The overall size is
decided by the character set and the feature vector size.
4.2.2. Output Layers
Output nodes are the set of all possible characters that are
present in license plate information. The number of out-
put nodes depends upon the numbers of characters that
need to be classified. Typically, it consists of 26 alpha-
bets “A - Z” and 10 numbers “0 - 9”. However, to reduce
algorithm complexity, certain countries avoid using both
0 and O together. In our test database, we have all the
characters except O, 1 and Q. Therefore, the number of
output layers is 33 in our implementation.
4.2.3. Hidden Layers
H id den layers are the intermed iate layer s between th e input
and output layers. There is no generic rule as such for
deciding the number of hidden layers. In our design, the
number of hidden layers was experimentally found to be 25.
For our implementation, we have used a supervised
learning paradigm. We obtained a set of 50 images for
each character from the database images and trained the
neural network using them. For better results, it is im-
portant to include images of all types of character repre-
sentation (ideal, noisy, see Figure 10) in the training
dataset. We have used MATLAB for training the neural
It was observed that for certain characters like 0-D and
5-S, error rate was higher compared to other characters,
due to similar spatial information content. In order to
improve the overall accuracy, a two stage detection pro-
cess for these characters is used. If the characters are
possibly between 5, 0, S and D, then region properties
like orientation, reflection coefficient were again used in
the 2nd stage identification process and the final possible
character was identified. This two stage process has im-
proved the overall algorithm accuracy by 1% - 2%.
5. Results
Performance evaluation for the license plate recognition
algorithm is a challenge in itself, since there is no common
reference po int for benchmarking [12]. We tested our algo-
rithm on 1500 different frames obtained from a sample vi-
deo, taken from commercial license plate recognition ca-
mera [14]. The resolution of all frames is 480 × 640. All
these frames were th en classified i nto following type s:
Type 1: In these samples, license plate is fully visible and
all its characters are clear and distinct, see Figure 11(a).
(a) (b)
(c) (d)
Figure 11. Different types of input frames (a) Type1 clear
and distinct; (b) Type 2 image with blurred license plate
details; (c) Type 3 license plate with dirt on it; (d) Type 4
presence of another text block in image; (e) Type 5 license
plate details in two rows.
Copyright © 2012 SciRes. JTTs
Type 2: Images in these samples are little blurred due
to variations in illumination conditions; see Figure 11(b).
Type 3: These samples have little dirt or shadows on or
near license plate region, Figure 11(c).
Type 4: In these samples, another text block apart from
license plate is present in the frame, Figure 11(d).
Type 5: License plate details are present in two rows,
Figure 11(e).
Figure 11 shows sample images for all the above im-
age types. Table 1 summarizes the license plate recogni-
tion results for all these image types.
A rare failure case is seen if the license plate is black
in color. In this case, the edge detection step fails since it
cannot locate license plate edges. One such case can be
seen in Figure 12.
6. Conclusions
The algorithm presented in this paper is extremely useful
in dealing with complex cases, while building a real-time
license plate recognition system. The use of Mexican hat
operator helps to improve the performance of edge dete-
ction step, when there is only partial contrast between li-
cense plate region and region surrounding it. Euler nu-
mber criterion for a binary image helps to reduce the
sensitivity of algorithm to license plate dimensions. Prep-
rocessing step using median filter and contrast enha-
Table 1. Results based on types of input images.
Type 1 879 99.4% 97.81%
Type 2 442 96.3% 96.63%
Type 3 153 94.77% 95.69%
Type 4 44 100% 99.25%
Type 5 11 100% 97%
Overall Results 1529 98.1% 97.05%*
*In the database of 335 vehicle images, each vehicle on an average has 5
frames. License of 330 vehicles were recognized with an accuracy of 98.5%.
(a) (b)
Figure 12. (a) Sample image with license plate black in color;
(b) Failure of edge detection step in locating license plate
ncement ensures performance in case of low. resolution
and blurred images. Local threshold operation prevents a
single brighter or darker region from corrupting the Thres-
hold value and thereby improves binarization process.
Reflection symmetry coefficient along with two stage
identification process improves the character recognition
performance significantly especially in dealing with com-
plex cases like recognizing 0 and D.
The algorithm can be extended to mobile license plate
recognition systems because of its ability to provide ex-
act license plate location in complex cases. Also, to incr-
ease the intelligence of license plate detection algorithm
while working with video input, motion analysis can be
applied to the sequential frames and selection of proper
frame can be done. The performance of algorithm will
also improve if higher resolution input frames are used.
[1] X. F. Shi, W. Z. Zhao and Y. H. Shen, “Automatic Li-
cense Plate Recognition System Based on Color Image
Processing,” International Conference on Computational
Science and Its Applications (ICCSA), Vol. 3483, 2005,
pp. 1159-1168.
[2] Z. Chen, C. Liu, F. Chang and J. Xu, “A Novel Algorithm
of License Plates Automatic Location Based on Texture
Feature,” IEEE International Conference on Automation
and Logistics, Shenyang, 5-7 August 2009, pp. 1360-1363.
[3] W. Pan and N. Yang, “A New Method of Vehicle License
Plate Location under Complex Scenes,” 2nd International
Conference on Advanced Computer Control (ICACC),
Shenyang, 27-29 March 2010, pp. 134-138.
[4] A. Delforouzi and M. Pooyan, “Efficient Farsi License
Plate Recognition,” 7th International Conference on In-
formation, Communication and Signal Processing, Macau,
8-10 December 2010, pp. 1-5.
[5] C. N. K. Babu and K. Nallaperumal, “A License Plate
Localization Using Morphology and Recognition,” IEEE
Indian Conference (INDCON), Kanpur, 11-13 December
2008, pp. 34-39. doi:10.1109/INDCON.2008.4768797
[6] B. M. Shan, “License Plate Character Segmentation and
Recognition Based on RBF Neural Network,” Second In-
ternational Workshop on Education Technology and Com-
puter Science (ETCS),Wuhan, 6-7 March 2010, pp. 86-89.
[7] H. Lan and G. Shrestha, “Real Time License Plate Reco-
gnition at the Time of Red Light Violation,” International
Conference on Computer Application and System Mo-
deling (ICCASM), Taiyuan, 22-24 October 2010, pp. 243-
[8] Y. Wen, Y. Lu, J. Q. Yan, Z. Y. Zhou, K. M. von Deneen,
and P. F. Shi, “An Algorithm for License Plate Reco-
gnition Applied to Intelligent Transportation System,”
IEEE Transactions on Intelligent Systems, Vol. 12, No. 3,
2011, pp. 1-16.
Copyright © 2012 SciRes. JTTs
Copyright © 2012 SciRes. JTTs
[9] J.-M. Guo and Y.-F. Liu, “License Plate Localization and
Character Segmentation with Feedback Self-Learning and
Hybrid Binarization Techniques,” IEEE Transactions on
Vehicular Technology, Vol. 57, No.3, 2008, pp. 1417-
1424. doi:10.1109/TVT.2007.909284
[10] Z.-X. Chen, C.-Y. Liu, F.-L. Chang and G.-Y. Wang, “Au-
tomatic License-Plate Location and Recognition Based on
Feature Salience,” IEEE Transactions on Vehicular Te-
chnology, Vol. 58, No.7, 2009, pp. 3781-3785.
[11] S.-L. Chang, L.-S. Chen, Y.-C. Chung and S.-W. Chen, “Au-
tomatic License Plate Recognition,” IEEE Transactions
on Intelligent Transportation Systems, Vol. 5, No. 1, 2004,
pp. 42-53. doi:10.1109/TITS.2004.825086
[12] C. Nikolaos, I. Anagnostopoulos, D. Psoroulas, V. Lou-
mos and E. Kayafas, “A License Plate-Recognition Algo-
rithm for Intelligent Transportation System Applications,”
IEEE Transactions on Intelligent Transportation Systems,
Vol. 7, No. 3, 2006, pp. 377-392.
[13] L. Keilthy, “ANPR System Performance,” Parking Trend
International, June 2008.
[14] “CitySync Single Lane PAL Demo Video,” 2011.
[15] R. C. Gonzalez and R. E. Woods, “Digital Image Proces-
sing,” 3rd Edition, Prentice Hall, Saddle River, 2008.
[16] X. Lin, J. Ji and Y. Gu, “The Euler Number Study of
Image and Its Applications,” 2nd IEEE Conference on In-
dustrial Electronics and Its Applications, Harbin, 23-25
May 2007, pp. 910-912.
[17] X. Wang, “Multiscale Median Filter for Image Denois-
ing,” IEEE 10th International Conference on Signal Pro-
cessing, Beijing, 24-28 December 2010, pp. 2617-2620.
[18] C.-Y. Ning, S.-F. Liu and M. Qu “Research on Removing
Noise in Medical Image Based on Median Filter Me-
thod,” IEEE International Symposium on IT in Medicine
and Education, Jinan, 14-16 August 2009, pp. 384-388.
[19] N. Otsu, “A Threshold Selection Method from Gray-
Level Histograms,” IEEE Transactions on Systems, Man
and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.
[20] Z. M. Li, K. P. Hou and H. Li, “Continuous Symmetry
Measure of Image,” IEEE International Conference on
Information Acquisition, Shandong, 20-23 August 2006,
pp. 824-828. doi:10.1109/ICIA.2006.305838
[21] A. K. Jain, J. C. Mao and K. M. Mohiuddin, “Artificial
Neural Networks: A Tutorial,” IEEE Computer Society,
Vol. 29, No. 3, 1996, pp. 31-44. doi:10.1109/2.485891