Automatic License Plate Detection System for Myanmar Vehicle License Plates

1
Khin Pa Pa Aung, 2Khin Htar Nwe, 3Atsuo Yoshitaka
1,2
University of Information Technology, Yangon, Myanmar
3
Japan Advance Institute of Science and Technology
1
khinpapaaung@uit.edu.mm, khinpapaaung.257@gmail.com
2
khinhtarnwe@uit.edu.mm, khinhtarnwe@gmail.com
3
ayoshi@jaist.ac.jp

Abstract
License Plate Detection is an intelligent system to find the is defined by the government of that country. In
exact license plate by analyzing image/video data for Myanmar, Vehicle registrations were started before 1996.
automatic license plate recognition (ALPR) system. The The current Myanmar plates are defined with three-letter
proposed system used image data. There is much early regional code (such as YGN for Yangon, MDY for
research has been done for ALPR purposes, however, it Mandalay, etc.) followed by six alphanumeric characters
was still challenging tasks for accurately detect license with the height of 16.5 cm and width of 36.2 cm [4].
plates in the open environment. The main difficulties lie in Moreover, Myanmar Vehicles Number Plates are not only
the diversity of plates such as language, font, color, and the state/division specific but also color specific for
type of the number plates that differ across nations and different types of services, such as black for personal
conditional variations such as various background scenes purposes, red for hire vehicles, yellow for religious, etc.
and illumination when captured. The fundamental ALPR Figure 1 shows some types of Myanmar License Plates.
system consists of three processes: license plate detection,
character segmentation, and character recognition.
Proposed system focused on license plate detection based
on image processing technology that is a crucial step for
the whole ALPR system. A Myanmar License plate has a
Figure 1. Different types of Myanmar license
white boundary for every different color of plates, thus the
plates
proposed system applied an edge-based approach for any
color of plates, and the plate region will remain because This study aims to obtain a specific and reliable
of its white boundary. After edge detection, license plate localization and recognition system for
morphological operation has been applied as it can add Myanmar’s license plates, mainly for dedicated parking
or remove pixels from/to the objects in an image. Thus the security.
license plate can be extracted accurately. Finally, the Normally, the ALPR system has three processes:
bounding box technology was applied to extract only the license plate detection, character segmentation, and
number plate region properly. character recognition. This work is only focused on
license plate detection based on image processing
Keywords- ALPR, license plate detection, character technology and it is important step for the entire ALPR
segmentation, diversity of plates, conditional variations system.
License plate detection is the technology for localizing
1. Introduction the region of interest from the given image/video only to
extract the desired location, that is, given input in this step
Automatic License Plate Recognition (ALPR) is a is vehicle image and the result will be extract number
technology to automatically recognize character string plate only. Although, a lot of research has been proposed
from number plate images attached to vehicles and it during the past years, till now it is a challenging task to
plays an essential trend in the field of the smart detect desired plate in the open situations such as blurring
transportation system. There are many potential the given image/video, capturing dissimilar angle
applications for the ALPR system such as security closed- conditions, the different aspect ratio of plates across
circuit television (CCTV), vehicle parking, traffic nations, etc. Overall plate detection algorithms can be
management and toll enforcement, and also many other classified into four different categories: character, texture,
applications. For each country, the license plate structure color, and edge based approaches [6].

132
 Character-based method is based on finding the character recognition from their different points of view
character like region in the image. Usually, in the vehicle and different types of challenges among researchers.
body, there is only the license plate will be the character Xiangjian He, Lihong Zheng [8] proposed vehicle plate
like region. However, nowadays most people used beauty extraction algorithm in which variable scanning window
stickers to decorate their vehicles. In such a case, moved to the whole input image and then using the well-
Character-based method will have remarkable trained classifier to classify either true decision that is
disadvantages. plate area or false decision that is not plate area. This
Color-based method is applied by finding out the work was the extension of P.Viola and M.J.Jones face
license plate colors that is normally different for each detection algorithm. The plate detection accuracy rate was
country. Also, this method focused on the color of the 96.4% and character segmentation was 98.82%. However,
vehicle. In [1] license plate is extracted based on HSI this work was not focused on skew correction of vehicle
color space model. This method can be applied to find the plates.
corrupted license plates. However, it is not suitable for the In [3], numbers from Myanmar License Plate were
same color and same size structure as a license plate in extracted by using neural network for Myanmar Vehicle
the given input image and very sensitive for conditionally Plate Recognition System. Plate localization is performed
illumination changes in nature. based on Euler number and aspect ratio, the resulted
Texture-based method is the pixel-based approach. accuracy for localization is 99%. However, the character
Compared with edge and color based methods it uses segmentation and recognition accuracy was 96% and
more discriminative characteristics, but it has 92%, respectively.
computational complexity for output. Jia Wang, Boris Bacic and Wei Qi Yan, 2017,
Proposed system applied an edge-based method. To proposed an effective method for plate number
find the license plate region in the given image edge- recognition. In this study license plate was localized by
based method identified edge density, edge shape, searching red light in HSV color space and then finding
boundary of the license plate, etc. This method is fast and vertical edge of a plate. The accuracy for localization and
simple. However, it requires finding the continuity of the recognition are 75% and 70%, respectively. This system
edges. Thus the system applied morphological operations: was focused on only back side vehicles’ plate and future
dilation, image filling, and erosion to remove unnecessary work was described for both sides of vehicle plate’s
edges and to extract connectivity of the objects. Lastly, recognition system [2].
the bounding box technology applied to extract only the O. Khin and S. Choomchuay, 2017, studied on
number plate region properly. Myanmar license plate detection only for dissimilar angle
vehicle plate. This research work totally 97% of number
Data Acquisition plates is well detected. However, this research work was
failed for car body color and plate color are the same, and
also for the license plate size is very small compared with
License Plate Detection the size of the vehicle [4]. Our system aims to propose a
reliable vehicle plate detection method and the result
indicated more than expected compared with the relevant
Character Segmentation contexts from our literature review.

3. Problem statements
Character Recognition
Nowadays, due to the number of automobiles,
Figure 2. Fundamental processes for ALPR monitoring vehicles for security purposes is consuming
system time, money and security officer to check the license plate
of every vehicle. It is not viable to employ several officers
The rest of the paper is organized as follows: a brief to work as whole-time license plate investigators. The
description of related work, problem statements, proposed only solution for that is the Automatic License Plate
system, experimental results, conclusion and future Recognition (ALPR) system and it will reduce the
research directions are elucidated in Section 2, 3, 4, 5 and workload and time for many security purposes.
6, respectively. The other three main factors why this problem was
taken up as a research topic are the facts: such as diversity
2. Related works of plates, environment variations, and very little research
work that has been performed over Myanmar License
A lot of the previous research has been done for Plates. First, the diversity of plates means that the location
license plate detection, character segmentation and and size of the license plate in each image frame are

133
different. In addition, plates may be tilted or obscured by given input image. Figure 4 shows the architecture of the
dirt, stickers, screws and other objects. Moreover, license proposed system.
plate structure is a country-specific. Thus the font,
language, color, and style of the license plates will differ
across nations. Therefore, the license plate recognition
system tends to be country-specific.
A second, conditional variation is depending on
various illuminations and background scenes: such as
input image/video may be captured at different lighting
conditions as well as the vehicle lighting. This is also
needed to be taken up into consideration. The
intricate/complicated background in the license plate:
such as plate-size advertisements, stickers, and other
writing numbers on vehicles that often give the false
alarm on the detection process.
In Myanmar, ALPR system is a current research topic. Figure 3. Collected images examples
It was still challenging tasks to accurately detect the
license plate in the open environment. Moreover, early Input Color Image
research has not been focused on regional codes
recognition: such as YGN, BGO, MDY, etc. Myanmar
vehicle plates have two rows, one for regional code and
Image Resizing
one for vehicle identifier as shown in Figure 1. Besides
the regional code, there has a round shape national seal
and screws that look likes character O and number zero
(0). It is also a challenge for the Myanmar license plate RGB to Gray Image
recognition system. In such a way, the ALPR system is Conversion Binarization
currently an interesting topic for many experimenters.

4. Proposed System Apply Edge Detector

4.1. Data acquisition

Apply Morphological
Data acquisition is the first step of the ALPR system. Operations
The outcome of the detection system also depends on the
quality of the images. This system tested 40 images
including both front and back sides of the vehicles. Most Apply Bounding Box Output Image
of these images are taken by using the phone camera from Algorithm Cropping
our University car parking. Images are taken from various (Output)
colors of vehicles for the proposed system. They are
saved with JPEG format since it is the default format of Figure 4. Architecture of proposed system
the phone camera. However, the system is working over
other formats such as PNG, etc. Figure 3 shows some The difficulty of color visualization and the
examples of collected images. complexity of color codes is the main reason we need to
convert color code in the image processing technique. The
4.2. Automatic license plate detection system conversion will reduce the complexity of the image. Thus
the system converts RGB to gray scale and as it is much
To provide an efficient and reliable license plate better for license plate detection compared with RGB
detection method proposed system applied image Image. Binarization is the next important process and it
preprocessing steps such as RGB to gray conversion, converts a pixel image to a binary image. Generally,
image binarization, image resizing. Then the system background and characters must be different colors in
applied sobel edge detector followed by some every license plate, so theoretically both have opposite
morphological operations: erosion, dilation, and image binary values. So, binary images are obviously easy to
filling. Lastly, the bounding box technology was applied extract license plate from input image. Imbinarize
to extract the license plate region accurately from the function converts the grayscale image by replacing value
1s for all values that are above the globally determined

134
threshold and replacing all the remaining values to 0s.
Imbinarize function uses 256-bin image histogram to
compute Otsu's threshold and this method also compute
by a specific method such as an adaptive method.
Proposed system utilized the former function, which
chooses the threshold value to minimize the intraclass (a)
variance of the thresholded black and white pixels.
As shown in Figure 1, Myanmar License Plate consists
of white font with different color background and
moreover, it has a white boundary on every plate. Thus
plate region obviously will remain after edge detection for (b)
any colors of car and any color of plates as shown in
Figure 5. Thus, the proposed system applied an edge-
based approach for license plate detection.
(c)

Figure 6. Morphological operation after edge

detection (a) Dilation (b) Erosion after dilation (c)
Plate extracted by bounding box

Figure 5. Output of Sobel detector

There are lots of edge detection methods and that has

been applied for different purposes. Prewitt operator does
not emphasize the pixels that are closer to the center of
the mask. Laplacian operator is very sensitive to noise and
to filter the noise we need to apply Gaussian smoothing
before the Laplacian filter. The canny edge detector is
also the most common and it is a complex edge detecting
method. In such a way, the system applied Sobel edge
detector. The Sobel edge enhancement filter has the
advantage of providing differentiating (which gives the Figure 7. Overall plate detection steps
edge response) and smoothing (which reduces noise)
concurrently. Figure 5 shows the output of the Sobel
operator. 5. Experimental results
Then, to know the exact place of the candidate license
plate in the input image, the system applied We experimented to verify the effectiveness of the
morphological operations: dilation, erosion, and image proposed method. Proposed method is implemented on
filling methods. The experimental performance accuracy Matlab2017 and tested over-collected images by the
was much satisfied. digital camera and from another source such as online
First, the system applied dilation to gradually enlarge images to test different formats of images. Altogether
the foreground boundary pixels (i.e. white pixels, there are 40 four-wheel vehicles images with JPEG and
typically) after the Sobel edge detector. Thus, the PNG formats. Almost all vehicle images are well-
foreground regions grow in size while holes within those detected. Figure 8 shows some of the detected images.
regions become smaller. Morphological dilation makes Only 1 out of 40 vehicles could not be detected
objects more visible and fills in small holes in objects as because of its degree orientation of the given input
shown in Figure 6(a). Then the system made erosion to images. This false rate is (40-39)/440 = 0.025. Hence the
remove islands and small objects so that only substantial accuracy rate is still high and gives optimal accuracy even
objects will remain as in Figure 6(b). After that, the taking this wrong detection of plate region. Although, the
system applied the bounding box to crop license plate system is not well detected the oriented input image, it
effectively as this method proves isolation between the has still many advantages. The system was detected both
background and object as shown in Figure 6(c). sides of the vehicle plates that are front and back sides see
Result of each of the plate detection steps is shown in in Figure 8. Also, it is well detected tilted or obscured
Figure 7 and it shows that the proposed method is good plates by dirt, stickers, screws and other objects as shown
enough compared with the relevant contexts of the license in Figure 8. Moreover, the proposed system can be
plate detection system. detected license plates that are captured at different

135
lighting conditions as shown in Figure 8(e) and Figure 7. Acknowledgement
8(f). Figure 8(e) was captured under the dark condition
and Figure 8(f) was captured at heavy lighting conditions. The author would like to say deep thanks to the
The remaining result of Figure 8 is under normal lighting honorable advisers for their calm guidance, keen
conditions. encouragement and valuable critiques for this research
work. Also, like to extend my thanks to the University of
Information Technology for offering me the require
resources in running this program. Finally, I pleased to
say thank my family for their great support and
inspiration in all respects of my study.
a
8. References
[1] K. Deb and K. Jo, “HSI color based vehicle license
plate detection", International Conference on Control,
b Automation and Systems, 2008 in COEX, Seoul, Korea.

[2] J. Wang, B. Bacic and W. Yan, “An effective method

for plate number recognition”, Multimed Tools Appl,
Springer Science+Business Media New York, 2017.
c
[3] Myint Myint Htay, “Localization and Recognition of a
Myanmar License Plate Based on Partially Cut Character
Structure”, Fourteenth International Conference on ICT
and Knowledge Engineering, 2016.
d [4] O. Khin, M. Phothisonothai and S. Choomchuay,
“License Plate Detection of Myanmar Vehicle Images
from Dissimilar Angle Conditions”, ICSEC 2017,
November 15-18, 2017.
e
[5] S. Du; M. Ibrahim; M. Shehata and W. Badawy
“Automatic License Plate Recognition (ALPR): A State-
of-the-Art Review”, IEEE Transactions On Circuits And
Systems For Video Technology, Vol. 23, No. 2, 2013.
f
[6] S. Du, M. Ibrahim, M. Shehata, and W. Badawy,
Figure 8. Samples of detected images “Automatic license plate recognition (alpr): A state-of-
the-art review," IEEE Trans. Circuits Syst. Video
6. Conclusion and future work Technol., vol. 23, no. 2, pp. 311-325, 2013.

As the license plate structure is defined by the [7] W. Zhou, H. Li, Y. Lu, and Q. Tian, “Principal visual
government of that country, it is not feasible to develop a Word discovery for automatic license plate detection,"
universal ALPR system. This paper proposed an efficient IEEE Trans. Image Process., vol. 21, no. 9, p. 4269-4279,
and reliable license plate detection system for Myanmar 2012.
Vehicle License Plates. This approach is well suited for [8] X. He, L. Zheng, Q. Wu, W. Jia, B. Samali and M.
the next step character segmentation that is in Palaniswami, “Segmentation of Characters on Car
progression. Future work will include the next two steps License Plates”, Multimedia Signal Processing (MMSP),
of the ALPR system that are character segmentation and 2008 IEEE.
character recognition with deep neural network.
According to the literature review, there is no previous [9] Hendry, R. Ching Chen “Automatic License Plate
research that has been focused on regional code Recognition via sliding-window darknet-YOLO deep
recognition such as YGN, MDY. Thus, regional code learning”, Image and Vision Computing Volume 87, July
segmentation and regional code recognition will focus in 2019, p. 47-56.
our future work.

136