Atlantis Highlights in Engineering (AHE), volume 1

International Conference on Science and Technology (ICST 2018)

Robust License Plate Detection using Convolutional

Neural Network
I Nyoman Gede Arya Astawa1, I Gusti Ngurah Bagus Caturbawa2, Elina Rudiastari3, I Made Ari Dwi Suta Atmaja4
Department of Electrical Engineering
Politeknik Negeri Bali
Bali, Indonesia
1
[email protected], [email protected], [email protected], [email protected]

Abstract—In general, the vehicle number plate detection datasets, but vehicle number plate detection in an open
system in an image should be able to overcome two problems, environment is still a challenging research (2). There are some
first how to determine the position of the vehicle number plate or difficulties in detecting license plates such as non-uniform
which vehicle number plate is on an image and the second how backgrounds, blur, inappropriate camera settings, shadows,
big is the plate. A number of vehicle license plate detection
and illuminations. These difficulties will often be found when
methods have been proposed over the past two decades, and some
have shown success in certain tasks. This study will detect the a license plate detection is implemented on mobile phone
license plate number of vehicles using the CNN. The initial devices.
process is to create a training data license plate numbers using In general, the vehicle number plate detection system in an
CNN processed on the server. Furthermore, the training data image should be able to overcome two problems, first how to
entered on the vehicle license plate detection applications. Vehicle determine the position of the vehicle number plate or which
license plate detection results will be displayed in percent vehicle number plate is on an image and the second how big is
accuracy. Based on the testing, the results obtained were very the plate. A number of vehicle license plate detection methods
satisfactory with the accuracy of detection of vehicle license have been proposed over the past two decades, and some have
numbers as much as 98.19%, the remaining 1.81% of the plates
shown success in certain tasks (3). The researcher (2) uses the
were detected but not the image of the plate that was cropped. It
shows that the method has been implemented on CNN that principal visual word (PVW) method by determining the
training data and license plate detection is very pretty accurate in position of the candidate character in the number plate
detecting vehicle license plate. Implementation of CNN methods identified for the first time, and the bounding box of the plate
on android based mobile devices for data processing training and is determined later and the result can automatically detect the
detection testing of vehicle license plate shows very satisfactory vehicle number plate. The researcher (1) models the function
results. that produces scores for each image sub-region using the
convolutional neural network (CNN) method which allows to
Keywords— license plate, detection, CNN, mobile device estimate the number plate location detected by several
vehicles on several highway lanes. CNN and Recurrent Neural
I. INTRODUCTION Network (RNN) methods are used by researchers (4), both
Mobile phone devices are not just limited to communication methods have advantages in learning features and joint image
devices. By using the application of computing technology, a or embedding labels. The results obtained are the performance
mobile phone device can perform different functions even at of these methods comparable to manual engineering methods
the same time. Implementation of the operating system in such as the sliding window method. The license plate
mobile devices make the device capable of computing detection study conducted by (5) based edge-geometrical
activities that are almost the same as computing devices in features using morphological approach. Researchers (3) used
general. This makes mobile devices possible to add the line density filter (LDF) method to extract candidate plate
applications based on artificial intelligence (AI) such as face areas, thereby significantly reducing the area to be analyzed
recognition, fingerprint recognition, vehicle license for plate localization. Researchers (6) using sliding window
recognition, etc. method, histogram of oriented gradients and support vector
License plates detection has a role in smart transportation machines for detection of vehicle license plate position on
systems. This can be applied in many applications such as android-based mobile phone get very satisfactory result.
speed measurement, vehicle recognition, estimated traffic Researchers (7) propose a plate detection algorithm based on
flow, security controls, automatic vehicle tickets, traffic saliency images applying a sliding window to identify the
monitoring etc (1, 2). There are several journals that have characters.
reported the results of license plate detection studies with very This study will detect the license plate number of vehicles
good results, even with almost perfect accuracy in their test using the CNN. The initial process is to create a training data

Copyright © 2018, the Authors. Published by Atlantis Press. 564

This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).
 Atlantis Highlights in Engineering (AHE), volume 1

license plate numbers using CNN processed on the server. data. The weights on the layer specify the convolution
Furthermore, the training data entered on the vehicle license kernel currently in use, so the convolution kernel can be
plate detection applications. Vehicle license plate detection trained based on the input on CNN.
results will be displayed in percent accuracy.
 Rectified Linear Unit (ReLU). Aims to apply the sigmoid
II. METHOD activation function to the activation output generated by
the previous layer.
A. Database Training
The first step in this research is to prepare 546 vehicle  Subsampling is the process of reducing the size of a data,
image that will be used as database training. Then start the according to (11) the use of a pooling layer on CNN only
cropping process on the vehicle number plate using the aims to reduce the size of the image so that it can be
labeling application. The crop image of the plate is stored in an easily replaced with a convolution layer with the same
.xml file. All crop files are collected and stored as .csv files. stride as the corresponding pooling layer.
Although the size of cropping results is not the same, this is not
a problem in the matching process. Unlike other methods that
 Fully Connected Layer is a layer commonly used in the
require the plate crop size to be exactly the same as the size of
the plate being tested. The next process is segmentation and application of Multi Linear Perceptron and aims to
extrusion feature using CNN method with TensorFlow library transform the data dimension so that data can be
and training result in the form of stored feature in the form of classified in a linear fashion.
inference or extension .pb. This process is shown in Fig. 1. C. License Plate Detection
Vehicle image Initially, the detection process starts with getting vehicle input.
Furthermore, the process of segmentation and feature
extraction. The result of the process goes into the matching
phase with the training database. The result of the process
Crop plate image produces vehicle plate image. The system diagram is
illustrated as shown in Fig. 2.

Segmentastion and Input vehicle

feature extraction image
with CNN

Segmentastion and
Save feature Database feature extraction
plate image training with CNN
a. Database
training
Fig. 1. Training process diagram.
Plate image

B. Convolutional Neural Network Fig. 2. License plate detection system diagram.

Convolutional Neural Network (CNN) is a very efficient
pattern recognition algorithm that is widely used in image III. IMPLEMENTATION
processing and pattern recognition (8). CNN has been well
applied for the detection, segmentation, and recognition of This vehicle number plate detection application is
implemented on Android-based mobile devices. Test data used
objects and regions in images such as the introduction of
in this research is to use vehicle image containing license plate
traffic alerts, biological image segmentation of face detection,
of Indonesia country vehicle, vehicle image taken using camera
text, pedestrian and human body (9). CNN has many of mobile device. There are 331 images of test data used. The
advantages such as simple structure, adaptability and few test data consists of the image of motorcycles and cars. Fig. 3
training parameters. CNN reduces the complexity of network shows an example of an application on a mobile phone.
models and the number of weights and has a unique advantage
in image processing and speech recognition (8). CNN has four Based on the testing, the results obtained were very
main operations (10), namely: satisfactory with the accuracy of detection of vehicle license
numbers as much as 98.19%, the remaining 1.81% of the plates
were detected but not the image of the plate that was cropped.
 Convolution Layer. The purpose of convolution on image
data is to extract features from the input image. The
convolution will produce a linear transformation of the
input data according to the spatial information in the

565
 Atlantis Highlights in Engineering (AHE), volume 1

TABLE I. RESULT OF VEHICLE PLATE DETECTION

Code Number of Plate Not Detected

Image Images Detected (Negative Image)
P1 39 39 0
P2 49 47 2
P3 56 55 1
P4 49 48 1
P5 54 53 1
P6 84 83 1
331 325 6

IV. CONCLUSION
Implementation of CNN methods on android based mobile
devices for data processing training and detection testing of
vehicle license plate shows very satisfactory results. The
accuracy of vehicle plate number detection is 98.19%. Our
research strongly recommends to use CNN method in the
continuation of research of character recognition of vehicle
license plate number.

ACKNOWLEDGMENT
Fig. 3. Plate detection with non-uniform lighting (a) Vehicle images (b) The authors would like to thank of Directorate General for
detecting plat position (c) cropping. Research strengthening and Development, Ministry of
Research, Technology and Higher Education, Republic of
It shows that the method has been implemented on CNN Indonesia as the sponsor of this research through the scheme of
that training data and license plate detection is very pretty prime university research grants.
accurate in detecting vehicle license plate. There is an increase
in results compared to previous studies conducted by sliding
REFERENCES
window method and histogram equalization (6). The results of
vehicle plate detection shown in table 1. Fig. 4 shows an
example of a missed detection on the vehicle. [1] F.D. Kurpiel, R. Minetto, and B.T. Nassu, "Convolutional Neural
Networks for License Plate Detection in Images," IEEE Proceed.
[International Conference on Image Processing (ICIP); 2017 17-20
September, 2017].
[2] W. Zhou, H. Li, Y. Lu, and Q. Tian, "Principal Visual Word Discovery
for Automatic License Plate Detection," IEEE Trans. on Image
Process., vol. 21, no. 9, pp. 4269 - 4279, 2012.
[3] Y. Yuan, W. Zou, Y. Zhao, X. Wang, X. Hu, and N. Komodakis, "A
Robust and Efficient Approach to License Plate Detection. IEEE Trans.
on Image Process., vol. 26, no. 3, pp. 1102-1114, 2017.
[4] T.K. Cheang, Y.S. Chong, and Y.H. Tay, "Segmentation-free Vehicle
License Plate Recognition using ConvNet-RNN," [International
Workshop on Advanced Image Technology, 2017].
[5] J.L. Tan, S.A.R. Abu-Bakar, M. M. Mokji, "License Plate Localization
Based on Edge-Geometrical Features Using Morphological Approach,"
IEEE Proceed. [International Conference on Image Processing,
September, 2013]
[6] I. Astawa, I.G.N.B. Caturbawa, I.M. Sajayasa, and I.M.A.D.S. Atmaja,
"Detection of License Plate Using Sliding Window, Histogram of
Oriented Gradient, and Support Vector Machines Method," J. Phys.
Conf. Series. vol. 953, no. 1, 2018.
[7] K.H. Lin, H. Tang, T. and S. Huang, "Robust License Plate Detection
Fig. 4. Result Detection of non-vehicle plate (a) vehicle image (b) result of Using Image Saliency," IEEE Proceed. [International Conference on
plate detection and negative image (c) Negative cropped images. Image Processing, September. 2010].
[8] T. Liu , S. Fang, Y. Zhao, P. Wang, J. Zhang, Implementation of
Training Convolutional Neural Networks. Cornell University Library:
USA, 2015, pp. 1-10.

566
 Atlantis Highlights in Engineering (AHE), volume 1

[9] Y. Lecun, Y. Bengio, and G. Hinton, "Deep Learning," Nature, vol. [11] J.T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller,
521, no. 7553, pp. 436-44, 2010. Striving for Simplicity: The All Convolutional Net, 2014.
[10] S. Das and J. Mukherjee, "Automatic License Plate Recognition
Technique using Convolutional Neural Network," Int. J. Comp. App.,
vol. 69, no. 4, pp. 32-6, 2017.

567