First International Conference on

Computer Science and Engineering

CCSE 2017
21-22 Feb. 2017
Faculty of Computer Engineering

License Plate Recognition: A Literature Survey

Mohammad Reza Pourghorban 1 Sajad Farokhi* 2 Mohammad Hossein Nadimi 3
Faculty of Computer Engineering, Faculty of Computer Engineering, Faculty of Computer Engineering,
Najafabad Branch, Islamic Azad Najafabad Branch, Islamic Azad Najafabad Branch, Islamic Azad
University, Najafabad, Iran University, Najafabad, Iran University, Najafabad, Iran
mr.pourghorban@sco.iaun.ac.ir fsajad2@pco.iaun.ac.ir nadimi@iaun.ac.ir

*Corresponding Author: fsajad2@pco.iaun.ac.ir (Sajad Farokhi)

Abstract— The license plate of the vehicle is the most appropriate method of authentication. Uncontrolled growth of the vehicle causes
problems in different fields including traffic control systems, highway management systems, electronic payment issues, parking
management strategies and the like. Automatic car plate recognition is defined as an automatic extraction of information from consecutive
images of license plates. In most analyzing systems, license plate recognition cameras capture images by recording black and white colors,
and in some cases, infrared cameras are utilized. It is worth mentioning that one of the main factors in the success of license plate
recognition systems is the high quality of the input images. In other words, the quality of images, attained in either outdoor or indoor
environmental conditions during daytime, is considerable. Nonetheless, to get a better result regarding recognizing license plate, a vast
range of methods has differently been presented and employed in various countries, cities, and even in states. Features of car plates for
different areas have usually distinctive background color, language characters, sizes and fonts. In spite of this fact, the license plate images
may be occluded with poor light conditions, soil, and blurring. License plate recognition systems usually come in three basic steps: 1) car
plate extraction; 2) characters segmentation; 3) characters recognition. In this paper, due to the limited number of review articles in this
field, a comprehensive review of the recent methods and techniques used in segmentation and recognition steps for license plate
recognition are presented. Herein, different articles have presented the latest approaches and according to the used techniques used as
well as based on segmentation and recognition steps in car plate recognition system, these investigations are set to be classified.
Additionally, comparison between different methods in terms of advantages, disadvantages, and recognition rate of output is done, and
finally, the current issues in this area are discussed so as to propose new challenges for future researches.

Keywords—Automatic license plate recognition, characters segmentation, feature extraction, classification

characters, and recognition characters. Considering the fact

I. INTRODUCTION that the weather in different regions has different atmospheric
According to current reports, more than half a billion cars in conditions and the major differences exist in the type of
the world are on the move, and each vehicle has an license plates, license plate locations, arrangement of the
identification number as a main identifier. Possessing an letters and numbers in license plates, the functional method
identification number or a license plate for each vehicle is an for car plate recognition is expected to be different for
obligatory legal norm; therefore, definitely, such procedure is different countries, as shown in "Fig. 1". In the first stage,
bound to encounter humans with some difficulties in terms of once an image of vehicle is captured, the extraction of car
controlling vehicles. Additionally, it is highly likely that this plate location in the original image is done. This mechanism
issue can be intensified if there are not any intelligent control is usually based on a set of local features such as boundary
systems. Based on this fact, the existence of systems for and color or edge detection.
recognizing license plate through cameras is highly vital. By The second stage, so-called “segmentation characters”, is
and large, using recognition systems are inevitably based on characteristics such as color, histogram, pattern
accompanied with some challenges. Among them, blurring matching algorithms and the like. Finally, recognition
[1], unfavorable weather conditions, resized images, rotated characters, based on template matching and feature extraction
images can be considered as outstanding impediments. as well as using classification such as neural networks, and
Commonly, when it comes to methods for organizing fuzzy, occur. The performance of the license plate
automatic license plate recognition systems, the three main recognition system depends on how strongly each phase of
stage are introduced [2], car plate extraction, segmentation the system functions.
 In this paper, we present a research on vehicle license plate images. An example of a rotated image on the paper and
recognition systems and methods which have been used in images for other cases such as noise and the non-uniform
recent years. Furthermore, we categories them into different illumination light are shown in "Fig. 2".
groups based on the approaches applied in segmentation and
recognition stages of the recognition procedure, and then
their advantages and disadvantages are compared so as to
open up some controversial topics for the future research.
The rest of this article is divided into the following parts: in
the Part II, categories associated with methods of characters
segmentation are presented. In the Part III, different ways for
recognition of license plate is studied. Finally, in the Part IV,
we give a brief overview and conclusions for future
researchers are drawn.

Fig. 1. Different car plates: (a) EGYPT car plate, (b) IRAN car plate [3, Fig. 2. Tilt correction and character segmentation under some defected
4]. conditions:(a) color breaking off, (b) contamination and fading, (c)
partial occlusion, and (d) two-layer license plate with contamination
and fading [8].
II. CHARACTER SEGMENTATION
The most important step in the automatic car plate This segmentation based on prior knowledge and the
recognition, after the plate was extracted from the original percentage of segmented analyses shows acceptable results.
image, is the characters segmentation. In this step, many Sarfaraz et al reported that [9] after the pre-processing of
problems and challenges exist including rotation, non-uniform the received image is done for overcoming relevant issues, in
illumination light, blurring, broken characters. Methods order to do segmentation, character segmentation is followed
provided for the license plate characters segmentation can be by using MBR on the binary image. In this way, each
divided into eight general categories [2, 5, 6]: character is based on bounding box segmentation. Eventually,
1) Using pixel connectivity the output of this stage is to identify the type of characters so
as to be sent to the next stage. The advantages of this method
2) Using projection profiles
can provide a platform in which the boundaries and areas
3) Using combined features depending on characters are not needed, whereas
4) Using prior knowledge of characters disadvantage of this method is the problem related to sticking
5) Using mathematical morphology characters. An example of this segmentation is shown in "Fig.
6) Using contours tracking and modeling 3".
7) Using histogram processing
8) Using local/global thresholding
According to Ktata et al [7], a method based on local
characteristics of pixels in the images of Tunisian license plate
was introduced. The proposed method for segmenting
characters was operated through neural network algorithms as
well as using local features and texture pixels in the grayscale
Fig. 3. Segmentation of license plate character: (a) binary map of detected
image in which the accuracy of pixel recognition rate of license plate, (b) character segmentation using minimum boundary
segmentation increases. The advantage of this method is its rectangle on connected components and bounding box [9].
high speed of segmentation characters, especially, when
images do not have any problems and challenges. However, if The same as other articles using this type of segmentation,
not detected properly due to problems, this method does not other researchers used bounding box for segmentation
perform well. character, and their approach for surmounting issues
Based on the work done by Wang [8], the car plates used concerning poor quality of images and dark image was
in China introduced challenges to the image, namely rotation successful [10, 11].
in different directions (vertical, horizontal) with different In another study done by Saeeb et al [3], they worked on
angles. Other challenges such as contamination, fading, images of license plate in Egypt and they proposed a different
partial occlusion are added to the input image, following that approach. In their investigation, the term “Egypt” is one of the
using the maximum and minimum local and methodology most considerable feature in the automatic license plate
presented in this article are set to improve the quality of
segmentation. As can be seen in "Fig. 4", this type of license
plate image is composed of three parts. The first part located
at the top of the plate is the word "Egypt" in both Arabic and
English writing styles, and the background color represents
the type of vehicle, including personal, taxi, military types.
Two other areas situated on the right and the left are letters
and numbers, respectively. This method is segmented in two
main stages; the first stage is based on recognizing the highest
area and its color needed for identifying the type of vehicle,
and in the second stage, the reduction of noise and removal of
unwanted elements are accomplished through using a low-
pass filtering so as to improve the quality. It also can remove
the dividing line between numbers and letters. Subsequently,
employing the binary image and the local feature image used
for segmentation of letters and numbers is performed.
Regarding the disadvantage of this method, it is highly likely
to ignore the challenges and problems such as blurring, weak Fig. 5. (a) Vertical edge maps of images, (b) Horizontal projection
intensity and fractured characters. histogram

Fig. 6. Binarized images of the images with upper and lower bounds
removed

Fig. 4. (a) The Egyptian plate, (b) Image with outer border, (c) Image
without border [3].

In a research done by S.H. Kasaei et al [4]. the importance

of having a pre-processing step and the importance of
segmentation stage were shown. Initially, they declared that
the image equipped with an improved filtering feature could
reduce unwanted items such as noise. Afterwards, the binary
image accompanied with a thresholding technique was used.
The threshold level was between 0.7 and 0.8. This technique
was empolyed since problems such as low quality, different
brightness and motion as well as bluring existed. In their Fig. 7. (a) The vertical projection of the images in Fig. 6, (b) Character
study, the study was followed based on morphological segments of license plates
operations for removing unwanted elements and then
characters were improved. Finally, by using the local mean of In a proposed method by Edson Cavalcanti Neto et al, the
each partition in the image, the character segmentation was length and width of the character as well as the distance of
accomplished. periods were used for identification. Their introduced method
was also based on the fact that the character spacing and size
In the trend provided by Zheng et al [12], they indicated were different, and their approach was adapted according to
that the segmentation, generally, was composed of three parts. different situations in which image filter was applied. After
In the first stage, using the high and low borders of characters, filtering, all characters located in different areas were
the structure of character such as its height could be achieved, detected, and they were defined by the rectangle. In this
and then segmentation characters could be done through using regards, red lines indicated the wrong area and green lines
the width of the character as well as employing Blob algorithm meant that the chosen area was correct, shown in "Fig. 8". The
and 8-connection with the pixels of binary image. Finally, area specified binary and could be resized during the daytime
segmented labels assigned to each character have not been and consequently it would be prepared for phase recognition.
changed in the process of identifying priorities for each An example of this process was shown in "Fig. 9" [13].
character. These mentioned three stages can be observed in
"Fig. 5", "Fig. 6" and "Fig. 7", correspondingly.
 In their research, neural network was used to classify
characters. This method was tested on 700 car plates and its
output was 90.78%. One of the disadvantages of this method
was about the fact that it did not consider pre-processing step
for improvement of other challenges.
Wang et al [8] declared that feature extraction could be
done by using (Scale Invariance Feature Transform) SIFT.
Such approach for overcoming the problems, including
adverse effects, complicated background, resizing, rotating to
different degrees, pollution, different lighting, partial
Fig. 8. Characters after filtering [13]. occlusion and defective characters was strongly
recommended. In their research, character recognition rate in
China was 96%. The disadvantage of this method can be
related to its sensitivity to noise and its low efficiency and
poor quality of images such as blurring can be noticed. An
example of this identification can be seen in "Fig. 11".
Another similar work was also done in which SIFT feature
extraction was used to recognize the license plate characters.
[1].

Fig. 9. Steps of segmentation Character: (a) Example of plate thresholding,

(b) All possible characters before filtering, (c) Characters after
filtering, (d) Result of characters segmentation [13].

III. CHARACTERS RECOGNITION AND CLASSIFICATION

Once the characters have been segmented, identification
system has entered a phase of recognition and classification of
the characters. Recognition of the characters of the license
plate in has a different method, depending on the previous
stages such as the first phase associated with the quality of the Fig. 10. A sample of Chinese character recognition: (a) SIFT feature points
output image, characters and segmentation-related problems of each template, (b) SIFT feature points of the potential target region
and challenges. Among the problems arising from this area, and (c) feature matching [8].
we can note resizing and the style of font characters used in
the car plate. Nonetheless, isolated characters may be noisy or In another research done by Sadeghi et al [2] on the Iranian
broken, and partial obstruction can be seen. Moreover, license plates, a normalization method based on neural
rotation and transformation could be observed in some cases. network and deliver images was used. In this way, several pre-
In the following list, different ways of recognition are shown processing on the image and then normalizing characters were
and it helps to classify groups of license plate characters, segmented. Additionally, through using a database of images
which are suitable for future use. Generally, vehicle license of characters and a neural network, the characters are
plate recognition methods can be divided into two categories: identified. The advantages of this method can be used to
recognize the shadows and angles with different views and the
1) Using raw data noisy images.
2) Using extraction features
According to the study done by Kasaei et al [4], templates
And it can also be presented and categorized based on the
matching different uses for the recognition of characters were
classification of algorithms into four general categories:
applied. For this purpose, the character size was fixed to the
1) K- Nearest Neighbor Classifier (KNN) dimensions of 38×20. This size was very important for
2) Naive Bayes Classifier matching the characters. This method was an effective
algorithm for recognition of the license plate characters. The
3) Neural Networks(NN) input characters were compared with all the database of
4) Support vector machine (SVM) characters and the best similarity was considered. The
Ktala et al used [7] local features extraction, average pixels disadvantage of this method was low efficiency in the
method, and recognition characters. This method can be implementation of Persian characters "‫ "ب‬and "‫"ت‬. This
associated with problems such as low-quality images, images approach was also used in other studies [3, 14] in which
containing both character disorders and other characters. template matching was used to identify characters.
 Based on the study carried out by Su et al [15], they used knowledge of the
a ring-to-line mapping for recognition of the license plate number of license
plate characters
characters was employed. Compared with other works, this
introduced procedure did not need character segmentation and Using prior Simple Limited by the prior
instead license plates were identified using Ring-to-line knowledge of knowledge, any
characters change may
mapping to make several Straight-line signals. The resulted result in errors
vector for transferring and rotating was highly applicable.
Based on this fact, this method could be used to remove the Using character Can get exact Slow and may
contours character boundaries generate incomplete
segmentation and it was outstandingly suitable as considering
or distorted
the rotation and the transferring. The disadvantage of this contour
method could be connected to the point that it ignored other
challenges if these stages as well as other challenges exist in Using combined More reliable Computationally
features complex
the input image, which resulted in reeducation of efficiency.
An example of this approach can be seen in "Fig. 12".

TABLE II. PROS AND CONS OF EACH CLASS OF LICENSE PLATE

RECOGNITION METHODS
Methods Pros Cons
Using pixel Template Simple and Processing
values matching comfortable nonimportant
pixels and
slow,
vulnerable
to any font
Fig. 11. Location (54,74) as centroid for extracting signals [15]. change,
rotation,
IV. DISCUSSION AND CONCLUSION noise
Several Be able to More
Segmentation and identification of vehicle license plate templates for recognize tilted processing
characters are two important steps in the automated license each character characters time.
plate recognition systems in which the efficiency and accurate Using Horizontal and Be able to Feature
identification depend on the types of used algorithm in the extracted vertical extract salient extraction
segmentation stage and recognition characters, and these two features projections features, robust takes
steps are interdependent. The number of
to any time, non
distortion, fast robust features
In recent years, several methods have been proposed in for black pixels in
recognition will degrade
these two steps, due to the limited number of review articles each 3 × 3 since the recognition.
in this area and also proposing layouts for the future pixels block the number of
features is
researching trends, this paper presented a comprehensive SIFT
smaller than
overview of methods for license plate characters that of the
segmentation, recognition and classification. Additionally, pixels.
various articles on positive and negative aspects as well as the
recognition rate of the characters have been compared. The
TABLE III. PERFORMANSCE COMPARISON SOME TYPICAL ALPR
result of this comparison showed merits and demerits of SYSTEMS [LICENSE PLATE SEGMENTATION (LPS), LICENSE
segmentation as shown in "Table. I". The results of PLATE RECOGNITION (LPR)]
comparison of recognition are shown in "Table. II" and results Methods Database Size LPS rate LPR rate
of comparison of recognition rates for segmentation and
recognition are shown in "Table. III". [7] 740 images 88.67% 90.78%
[8] - 86% 96%
TABLE I. PROS AND CONS OF EACH CLASS OF LICENSE PLATE
SEGMENTATION METHODS [9] 12 s video 88% 93%
Methods Pros Cons [3] 100 images 93% 91%
Using pixel Simple and Fails to extract all the [4] 150 images 94% 92%
connectivity Comfortable, robust characters when there
to the license plate are [12] - 98.82% 94%
rotation joined or broken [13] 700 videos 99.65% 98.5%
characters
[10] 1000 Images 98% 98.2%
Using projection Independent of Noise affects the
profiles character positions, projection value, [11] - 97.6% 90.1%
be able to deal with requires prior
some rotation [2] 120 images 98% 91%
[14] - 98.33% -
 [15] 120 images - 93% recognition for CCTV forensic applications," Journal of real-
time image processing, vol. 8, pp. 285-295, 2013.
[10] Shapiro, V. Gluhchev, G. and Dimov, D. "Towards a
As long as the license plate recognition systems in today's multinational car license plate recognition system," Machine
society are vital, researches on existing problems as well as Vision and Applications, vol. 17, pp. 173-183, 2006.
finding solutions for improvement of these systems are very [11] Jiao, J. Ye, Q. and Huang, Q. "A configurable method
crucial. for multi-style license plate recognition," Pattern
Recognition, vol. 42, pp. 358-369, 2009.
The following topics for future researches in this area can [12] Zheng, L. He, X. Samali, B. and Yang, L.T. "An
be outlined: algorithm for accuracy enhancement of license plate
1) To tackle with the differences in the intensity of recognition," Journal of Computer and System Sciences, vol.
light and shadow in a picture, working on 79, pp. 245-255, 2013.
methods so as to improve the image pre- [13] Neto, E.C. Gomes, S.L. Rebouças Filho, P.P. and de
processing, and eliminating the problems and Albuquerque, V.H.C. "Brazilian vehicle identification using
challenges of light. a new embedded plate recognition system," Measurement,
2) Improving the accuracy of identifying the vol. 70, pp. 36-46, 2015.
problems caused by difficulties regarding [14] Gazcón, N.F. Chesñevar, C.I. and Castro, S.M.
distinguishing characters such as (B, 8), (O, 0), "Automatic vehicle identification for Argentinean license
(I, 1), (A, 4), (C, G), (D, O), (K, X) as well as plates using intelligent template matching," Pattern
broken characters. Recognition Letters, vol. 33, pp. 1066-1074, 2012.
[15] Su, C.-L. "Car plate recognition by whole 2-D image,"
3) Overcoming problems and researching on Expert Systems with Applications, vol. 38, pp. 7195-7200,
blurred images and improving the quality of 2011.
images in the pre-processing stage.
REFERENCES
[1] Chen, H. Hu, B. Yang, X. Yu, M. and Chen, J. "Chinese
character recognition for LPR application," Optik-
International Journal for Light and Electron Optics, vol. 125,
pp. 5295-5302, 2014.
[2] Sedighi, A. and Vafadust, M. "A new and robust method
for character segmentation and recognition in license plate
images," Expert Systems with Applications, vol. 38, pp.
13497-13504, 2011.
[3] Massoud, M. Sabee, M. Gergais, M. and Bakhit, R.
"Automated new license plate recognition in Egypt,"
Alexandria Engineering Journal, vol. 52, pp. 319-326, 2013.
[4] Kasaei, S.H. Kasaei, S.M. and Kasaei, S.A. "New
Morphology-Based Method for RobustIranian Car Plate
Detection and Recognition," International Journal of
Computer Theory and Engineering, vol. 2, p. 264, 2010.
[5] Zaitoun, N.M. and Aqel, M.J. "Survey on Image
Segmentation Techniques," Procedia Computer Science, vol.
65, pp. 797-806, 2015.
[6] Panchal, T. Patel, H. and Panchal, A. "License Plate
Detection Using Harris Corner and Character Segmentation
by Integrated Approach from an Image," Procedia Computer
Science, vol. 79, pp. 419-425, 2016.
[7] Ktata, S. Khadhraoui, T. Benzarti, F. and Amiri, H.
"Tunisian License Plate Number Recognition," Procedia
Computer Science, vol. 73, pp. 312-319, 2015.
[8] Wang, Y. Ban, X. Chen, J. Hu, B. and Yang, X. "License
plate recognition based on SIFT feature," Optik-International
Journal for Light and Electron Optics, vol. 126, pp. 2895-
2901, 2015.
[9] Sarfraz, M.S. Shahzad, A. Elahi, M.A. Fraz, M. Zafar, I.
and Edirisinghe, E.A. "Real-time automatic license plate