You are on page 1of 5

License Plate Localization based on Kapur Optimal

Multilevel Threshold

Nur Aliyatul Husna Bt Yahya Siti Norul Huda Bt Sheikh Abdullah


Center for Artificial Intelligence Technology Center for Artificial Intelligence Technology
Faculty of Information Science and Technology, Faculty of Information Science and Technology,
The National University of Malaysia The National University of Malaysia
Selangor, Malaysia Selangor, Malaysia
aliyatulhusna@gmail.com snhsabdullah@ukm.edu.my

Abbas Salimi Bin Zaini Mohd Zamri Murah


Center for Artificial Intelligence Technology Center for Artificial Intelligence Technology
Faculty of Information Science and Technology, Faculty of Information Science and Technology,
The National University of Malaysia The National University of Malaysia
Selangor, Malaysia Selangor, Malaysia
zabbassalimi@gmail.com mohdzamrimurah@gmail.com

Azizi Bin Abdullah


Center for Artificial Intelligence Technology Shariffpudin Basiron
Faculty of Information Science and Technology, Synergy Saver (M) Sdn Bhd,
The National University of Malaysia Selangor, Malaysia
Selangor, Malaysia shariff@my-synergy.com
azizia@ukm.edu.my

Abstract— A license plate localization system is useful for Speeded Up Robust Features (SURF) and Bag-of Words
many applications. Due to ambient of lighting in three distinct (BoW) feature descriptors. They were combined and clustered
situation which are morning, afternoon and night causing by using K-means clustering to form a novel way of localizing
difficulty to search optimum threshold value in each situation. the license plate’s region in an image. Makaoui et al. [4]
This research uses global thresholding approach by using Kapur inspect the car vertical edges to locate the plate by using Haar
entropy multilevel threshold based on Patch-Levy Bees wavelet method with several edge detection algorithms such as
Algorithm (PLBA). As a result, the system properly localize and Sobel and Canny algorithms, as well as morphological,
identify number plate in the image by using proposed including the black-hat transform. Abidin[2] proposed
segmentation image. From the experiment, proposed method are
multilevel thresholding method based on combination of
achieve accuracy rates to 67.68%, 90.71%, 24.34% respectively
for morning, afternoon and night dataset.
maximum entropy. However, performance of the method gives
a low level in segmentation process because the chosen
Keywords—Image segmentation; Multilevel thresholding; threshold value does not apply optimization process. Abdullah
License Plate Localization et al. [1] proposed adaptive thresholding by finding peak of the
blob distributions to identify the license plate number of the
vehicle due to ambient of lighting. Wen et al. [5] proposed high
I. INTRODUCTION dynamic range (HDR) method to enhance the image and the
Researches on license plate recognition are growing to morphology operator is applied to locate the license plate in the
become more robust with the existing proposed methods. Some image followed by binarization, tilt correction, and
of the important applications are solid access control, segmentation to obtain several upright license characters.
systematic traffic control, automation of toll collection, Reseachers shows that even though there are many usage of
identification of stolen vehicle, border control and law multilevel threshold in recent studies, some of the methods are
enforcements [1,2]. License plate recognition generally still having problems in robustness and computationally
comprised several steps that are image capturing, image intensive [1][7]. Hence, multilevel threshold can be improve
localization, feature extraction and classification. In this paper, further.
the proposed method is based on image localization method.
In Malaysia, there are some difficulties for the system to
There are various studies that have been done in obtain a good localization result. According to Synergy Saver
localization method. Khaleel [3] proposed localization by using CEO Mr. Shariffpudin Basiron, their company have tested

978-1-5090-3519-9/17/$31.00 2017
c IEEE 77
several LPR products and failed to achieve high accuracy rate binary image from gray level image into black and white
because of unstandardized license plate format practice due to image by turning all the pixels below some threshold value to
the lack of enforcement in Malaysia. Some of the license plates zero and all the pixels that upper the threshold value into one.
are in the form of a single or a double line with normal fonts. In this paper, multilevel thresholding scheme for Kapur’s
Others are made of special fonts, such as italic-like fonts as in based on the Patch-Levy-based Bees Algorithm (PLBA) is
[1]. Fig. 1 shows samples of license plates in Malaysia. proposed for license plate localization. In this section, Kapur’s
objective functions are presented and then the multilevel
thresholding PLBA-based algorithm is explain for searching
the optimal thresholds based on Kapur’s objective functions.
Assume that an image can be represented by L gray levels.
The probability of gray level i in the image can be calculated
as:
Fig. 1. Samples of license plates in Malaysia

(1)

Where,

h(i) = number of pixel,I in the image

= total pixel in the image

If the image divide into three classes, A, B and C based on


number of threshold t,where class A containing gray levels
from 1 until t1, B containing gray levels t1+1 until t2 and C
containing gray levels t2+1 until L normalization probability
pA, pB and pC are as follow:

Fig. 2. License Plate Localization (2)

There are many obstacles to get accurate result. Locating an


object in an image can be a difficult problem due to complexity (3)
of image background license plate, vehicle color and variety
type, and change as a result of different weather conditions and
other factors in different light, making accurate positioning (4)
difficulty [6]. The proposed localization method consists of
three steps; image segmentation, blob agglomeration and Where,
location of license plate as in Fig 2. To improve the
localization method we are include the segmentation process to
subdivide image into several classes by using two types of
maximum entropy which is Kapur entropy based on Patch
Levy Bees Algorithm. Hussein et al. [7] introduced Patch Levy
Bees Algorithm based on Kapur entropy to investigate the (5)
ability of the method to search optimum multilevel threshold
values. This is due to the increasingly level of thresholds will
extend the computational time. To reduce the computational
(6)
time, metaheuristic search algorithm helps to find optimum
threshold in a reasonable time [8].

II. MOTIVATION (7)

A. Multilevel thresholding
One of the popular technique in image segmentation is
thresholding method. Thresholding process is to generate

78 2017 7th International Conference on Cloud Computing, Data Science & Engineering – Confluence
Kapur Entropy for the class A, B and C are:

(8)

(9)

(10)

The entropy of threshold image is calculated by


following objective function:

H(t) = EA(t) + EB(t) + EC(t) (11)

Kapur entropy method selects the optimal threshold


by chosen maximum entropy as bellow:

t*= arg max(H(t)) (12) Fig. 3. Summarize of PLBA method [8]

B. Patch Levy Bees Algorithm (PLBA) III. METHODOLOGY


Patch and Levy are to enhance the initialization stage of Research method contain three steps which are image
basic Bees algorithm in local and global parts. A Levy flights is segmentation using PLBA-Kapur, blob agglomeration and
a random walk in which the step-lengths have a probability location of license plate. Steps are described as below.
distribution. Generating step-lengths based to Levy distribution
is advantageous for local parts as short steps in making more A. Multilevel thresholding
population members exploit the more convincing areas. At the
same time, the rarely long jumps happened which help keeping The first step in localization process is image segmentation
the members portion to explore the distant areas of solution to find optimum threshold value by using PLBA-Kapur. After
space as the advantageous for global parts. As a result, it is multilevel threshold is discovered, the maximum threshold
sprinting the convergence to the optimal solutions. After the value between multilevel values is applied to obtain the
initialization stage, the evaluation of the population is takes number plate. This is due to the number plate is in the
place. If the criterion for stopping is not met, a new population maximum intensity of gray scale. In this stage, the images is in
will be generated by selecting other sites for neighborhood binary image that containing significant object in the form of
search. A local search algorithm called greedy local search white blobs. However, white blobs are also containing noise.
based on Levy flight works at the same time reduces the length
of the long steps of Levy flight and frequent short steps on B. Blob Agglomeration
exploiting the convincing areas, hence, new optima can be
found. The process continue with the remaining bees After the threshold process, contour is applied to find
redistribute inside the hive to the patches according to Levy significant object (blobs of white pixels) on the image using
flight distribution to scout again and evaluate their fitness until basic geometric equations. Each potential blobs are enclosed
the criterion is met [8,9]. The PLBA method summarize as with rectangle label. Blobs that not differ much based on their
below: vertical coordinate, distance and height are clustered. The
selected blob is considered as the winner blob.

2017 7th International Conference on Cloud Computing, Data Science & Engineering – Confluence 79
C. Location of License Plate

For the localization part, after the winner blob is identified,


the blobs are compared with real character of the number plate.
If the character is correctly classify, then the localization is
successful. The accuracy of the localization is calculated based
on the score of the hits character in the localization.

IV. EXPERIMENTAL RESULT


The image collection (dataset) was obtained from Sistem
Lingkaran-Lebuh Raya Kajang (SILK) Highway-Sungai Long
Toll Plaza, a highway in Malaysia. The dataset are divided into
3 categories based on morning (297 images), afternoon (140
images) and night (152 images).

Fig. 5. Accuracy of [1] with PLBA-Kapur for morning dataset

(a) (b) (c) Table II shows the accuracy of multilevel threshold based
on [1] with PLBA - Kapur which is proposed method
Fig. 4. Sample of UKM-Synergy dataset in the (a) morning, (b) afternoon compared with adaptive threshold [1] using Kittler and
and (c) night Illingworth method and SURFBoW [3] using Kittler and
Illingworth thresholding method. The accuracy for the methods
The proposed method was tested to LPR system. The
show that afternoon dataset is achieved higher accuracy;
results obtained by proposed multilevel threshold is compared
90.71%, 80.71% and 85% correspondingly. This is due to the
with adaptive threshold [1] using Kittler and Illingworth
medium high contrast that helps to reduce the illumination
method and SURFBoW [3] using Kittler and Illingworth
problems. However, the system deals with different of
thresholding method. Table I and fig. 4 and 5 show the
illumination that is irregular contrast in morning dataset images
accuracy for morning, afternoon and night dataset by using
gives the percentage 67.68% accuracy of [1] with PLBA -
PLBA-Kapur with different number of threshold. For morning
Kapur, 62.29% accuracy of [1] using Kittler and Illingworth
dataset show that, two threshold is the higher accuracy
thresholding method and 58.25% accuracy of [3] using Kittler
(67.68%). For afternoon dataset, the higher accuracy (90.71%)
and Illingworth thresholding method. For night dataset, which
at two number of threshold among the dataset which give
has very low contrast of lighting causes the accuracy is very
sufficient of lighting compared to morning and night dataset.
low. Accuracy of [1] with PLBA - Kapur, [1] using Kittler and
Night dataset also give high accuracy at two number of
Illingworth thresholding method and [3] using Kittler and
threshold which is 24.34%.
Illingworth thresholding method for night dataset is 24.34%,
40.79% and 56.58%. Reflections of car lighting towards
TABLE I. ACCURACY FOR MORNING, AFTERNOON AND NIGHT DATASET cameras also may causes the license plate was hidden and
BY USING [1] WITH PLBA-KAPUR WITH DIFFERENT NO.OF THRESHOLD effected the accuracy of localization.
Morning Dataset The accuracy of [1] with PLBA - Kapur is higher than [1]
No of threshold 2 3 4
using Kittler and Illingworth thresholding method and [3]
using Kittler and Illingworth thresholding method for morning
Accuracy of [1] with PLBA -
Kapur
67.68% 57.23% 52.18% and afternoon dataset because the ability of the optimization of
threshold to search the optimum classes of the multilevel
Afternoon Dataset
images. The competent object can be obtained in the image
No of threshold 2 3 4 during the daylight. However, [3] using Kittler and
Accuracy of [1] with PLBA - Illingworth thresholding method is having higher accuracy at
90.71% 87.86% 83.57% night. Thus, [1] with PLBA - Kapur can be improved.
Kapur
Night Dataset

No of threshold 2 3 4
Accuracy of [1] with PLBA -
24.34% 15.13% 9.21%
Kapur

80 2017 7th International Conference on Cloud Computing, Data Science & Engineering – Confluence
Illingworth thresholding. Experimental results show that the
Sample Original Image Segmentation image proposed multilevel threshold works better on daylight dataset
Dataset and improved the localization accuracy. However, the
a)Morning
proposed method can be further improved to increase the
accuracy in the future.

ACKNOWLEDGMENT
This project is funded by PRGS/1/2016/ICT02/UKM/02/1
b)Afternoon GRANT entitled “Intelligent Vehicle Identity Recognition for
Surveillance”, DIP 2015-023 entitled “ Object Descriptor via
Optimized Unsupervised Learning Approaches” and
ERGS/1/2013/ICT01/UKM/03/5 entitled “Pembelajaran
Separa Selia Bagi Penggolongan Kata Jawi Untuk Manuskrip
c)Night Melayu” are used to fund this project.

REFERENCES

[1] Abdullah, S. N. H. S., Omar, K., Zaini, A. S., Petrou, M. and Khalid,
M, “Determining adaptive thresholds for image segmentation for a
license plate recognition system,” International Journal of Advanced
Computer Science & Applications, vol. 1(7), pp. 510-523. 2016.
[2] Abidin, N. H. Z., Abdullah, S. N. H. S., Sahran, S. and PirahanSiah, F.
Fig. 6. Sample segmentation image in the (a) morning, (b) afternoon and (c) 2011. “License plate recognition with multi-threshold based on
night entropy,” International Conference on Electrical Engineering and
Informatics (ICEEI), pp. 1-6, 2011.
[3] Khaleel, F. M., Abdullah, S. N. H. S. and Ismail, M. K. B., “License
ACCURACY OF [1] WITH PLBA-KAPUR, [1] WITH KITTLER AND ILLINGWORTH plate detection based on Speeded Up Robust Features and Bag of Words
AND [3] WITH KITTLER AND ILLINGWORTH THRESHOLDING BASED ON model,” IEEE International Conference on Smart Instrumentation,
MORNING, AFTERNOON AND NIGHT DATASET FROM UKM-SYNERGY. Measurement and Applications (ICSIMA), pp. 1-5. 2013.
Morning Afternoon Night [4] Makaoui, K., Guennoun, Z. and Ghogho, M., “Improved license plate
Dataset Dataset Dataset localization,” International Conference on Electrical and Information
a) Accuracy of [1] with Technologies (ICEIT), pp. 402-405. 2016.
67.68% 90.71% 24.34%
PLBA - Kapur [5] Wen, C. Y., Huang, T. S., Tseng, C. C. and Huang, S. S., “License
b) Accuracy of [1] with plate localization and recognition under different illumination
62.29% 80.71% 40.79%
Kittler and Illingworth conditions,” IEEE International Conference on Consumer Electronics-
c) Accuracy of [3] with Taiwan (ICCE-TW), pp. 1-2. 2016.
58.25% 85% 56.58%
Kittler and Illingworth [6] Abdullah, S. N. H. S., Sudin, M. N., Prabuwono, A. S. and Mantoro,
T., “License plate detection and segmentation using cluster run length
smoothing algorithm,” Journal of Information Technology Research
(JITR), vol. 5(3): 46-70. 2012.
V. CONCLUSION
[7] Hussein, W. A., Sahran, S. and Abdullah, S. N. H. S, “Patch-Levy-
In this paper, multi-level threshold based on PLBA-Kapur has based initialization algorithm for Bees Algorithm,” Applied Soft
Computing 23, pp.104-121, 2014.
been proposed for Malaysian License Plate Localization. This
[8] Hussein, W. A., Sahran, S. and Abdullah, S. N. H. S., “A fast scheme
is due to the variations of ambient lighting that interrupt the for multilevel thresholding based on a modified bees algorithm,”
system to localize the license plate correctly. The method has Knowledge-Based Systems, vol. 101, pp.114-134, 2016.
been compared with adaptive threshold [1] by using Kittler [9] Gutowski, M., “Levy flights as an underlying mechanism for global
and Illingworth and SURFBoW [3] with Kittler and optimization algorithms”, 2001.

2017 7th International Conference on Cloud Computing, Data Science & Engineering – Confluence 81

You might also like