Malaysian car plate localization using region-based convolutional neural network

Bulletin of Electrical Engineering and Informatics
Vol. 9, No. 1, February 2020,pp.411~419
ISSN: 2302-9285,DOI: 10.11591/eei.v9i1.1862  411
Journal homepage: http://beei.org
Malaysian car plate localization using region-based
convolutional neural network
Tay Eng Liang1,Usman Ullah Sheikh2, Mohd Norzali Haji Mohd3
1
Intel Microelectronics (M) Sdn. Bhd., Malaysia
2
School of Electrical Engineering, Faculty of Engineering, Universiti Teknologi Malaysia, Malaysia
3
Faculty of Electrical and Electronics Engineering, Universiti Tun Hussein Onn Malaysia, Malaysia
Article Info ABSTRACT
Article history:
Received Sep 3, 2019
Revised Nov 11, 2019
Accepted Dec 18, 2019
Automatic car plate localization and recognition system is a system that
identifies the car plate location and recognizes the characters on the car plate
input images. Within the automated system, the car plate localization stage
is the first stage and is the most crucial stage as the success rate of the whole
system depends heavily on it. In this paper, a Malaysian car plate localization
system using Region-based Convolutional Neural Network (R-CNN)
is proposed. Using transfer learning on the AlexNet CNN, the localization
was greatly improved achieving best precision and recall rate of 95.19%
and 97.84% respectively. Besides, the proposed R-CNN was able to localize
car plates in complex scenarios such as under occlusion.
Keywords:
Car plate localization
Deep learning
R-CNN
Transfer learning
This is an open access article under the CC BY-SA license.
Corresponding Author:
Usman Ullah Sheikh,
School of Electrical Engineering, Faculty of Engineering,
Universiti Teknologi Malaysia,JohorBahru, Malaysia.
Email: usman@fke.utm.my
1. INTRODUCTION
Automated car plate localization and recognition system is one component of an intelligent
transportation system. It uses image-processing techniques to automatically identify car plates from given
input images without any human intervention. This system has a very wide application area which includes
traffic monitoring systems,car park access systems,traffic law enforcement systems,automatic toll payment
systems, and border crossing control system [1-3]. In order to achieve a good car plate recognition result,
the system must first be able to accurately localize the location of the car plate from the given input image.
In general, there are two major branches of algorithms being applied to perform the automatic car plate
localization task, namely algorithms based on handcrafted features and algorithms based on deep learning.
Example of handcrafted features includes the location of straight lines, edge density, connectivity
information,and color information.Besides, in Malaysia,there are multiple car plate formats available [4, 5].
Consequently, methods based on handcrafted features will have difficulty in handling the issue
of non-standardization of Malaysian carplate formats.
In [6], the authors have proposed to localize car plate from an input image by applying Hough
Transform in order to detect the location of straight lines in the image and retrieve the bounding box location
based on the location of the straight lines. However, Hough Transform requires large memory and has high
computation time. As for [7-9], the Sobel operator has been used to localize car plates based on vertical edge
detection. The limitations of vertical edge detection are that it requires prior knowledge about the edge
information in the image and it is sensitive to noise. Authors in [10] have proposed to localize car plates

 ISSN: 2302-9285
Bulletin of Electr Eng & Inf, Vol. 9, No. 1, February 2020 : 411 – 419
412
based on connected components analysis (CCA) while authors in [11] have proposed to use image
binarization with the Otsu method. CCA may generate broken objects while the Otsu method is less accurate
when both car body and carplate have the same color.
Deep learning approach, or also known as deep structured learning or hierarchical learning [12],
replaces the need for handcrafted features [13]as the process of determining and learning features is achieved
by the algorithm through a large number of training images [14]. There are various deep learning
architectures available, which include Convolutional Neural Network (CNN), Deep Belief Network (DBF)
and Deep Stacking Network (DSN) [12]. Each architecture has its own application area, for instance, CNN
is suitable for object recognition [15]. Previous literatures have applied deep learning algorithm to localize
car plates [14, 16-18] using CNN but have some limitations such as high false negative rate as low-resolution
dataset has been used in the training process, assumption that there is only one car plate per image and they
did not consider the complexity of Malaysian car plates, therefore, not suitable to be applied for Malaysian
car plate localization system. Hence, a Malaysian car plate localization system using Region -based
Convolutional Neural Network (R-CNN) is proposed. By using this deep learning approach, the dependency
on handcrafted featurescan be eliminated and the issue of non-standardization of Malaysian carplate formats
can be handled by the proposed system. The proposed system also aims to be able to localize all car plates
from an input image if there are multiple car plates exists in the image. The rest of this paper is organized
as follows. Section II describes the proposed algorithm, tools used and evaluation criteria used to evaluate
the developed system. Results, analysis, and discussions are presented in Section III. Finally, the conclusion
and recommendationsforfuture work are discussed in Section IV.
2. CAR PLATE LOCALIZATION WITH R-CNN
The scope of this work is on Malaysian car plates with a black background and white characters
for normal cars as well as a white background and black characters for taxies. The circumstances that are
considered include multiple car plates (up to 3 car plates per image), special car plates (Putraj aya,
Perodua,etc.), car plateswith other logos, partially blocked car plates and carplates captured at an angle.
2.1. Image acquisition and preprocessing
The image acquisition process is carried out in order to collect the input dataset required to train
and test the proposed system. In this work, a total of 1.157 Malaysian car plate images were acquired
consisting of front and rear car images at 13 Megapixel resolutions in normal daylight condition. Each car
plate image consists of a minimum of 1 car plate and up to a maximum of 3 car plates. The car plates
in the input image take up approximately 20% to 30% of the whole image area. Images were also captured
at different angles with respect to the camera with a camera angle of between 0° to 30°. Images are first
converted to grayscale as the Malaysian car-plate consists of only black and white colors. Next, images
are resized to 826×465 from the original resolution 4128×2322 - to reduce compute time while maintaining
car plate image quality. After that, median filter is applied to the grayscale images in order to remove image
noise [19] before applying image intensity adjustment to enhance the contrast.
2.2. Training the AlexNet
AlexNet is a pre-trained CNN network that has been created by Alex Krizhevsky et al. [20].
It is an 8 layer CNN model that has been trained with a total of 1.2 million training images obtained from
ImageNet Large Scale Visual Recognition Challenge 2010 (ILSVRC-2010) with 1.000 object classes.
R-CNN is a network that combines CNN network with region proposals [21], which has been proven to be
successful for semantic segmentation [22] and object detection [23]. In order to perform transfer learning
based on AlexNet, the last few layers of the AlexNet are replaced with the layers that are applicable
for the current task. The last fully connected layer is changed to achieve two output classes, which are; car
plate class and the background class. Transfer learning is then carried out in 3 iterations. In the first iteration,
ground-truths are labeled by enclosing only the car plate numbers. The purpose of doing so is to allow
the R-CNN network to learn the car plate characters. Figure 1(a) shows an example of labeled ground-truth
for transfer learning in iteration 1. In this iteration, a total of 693 labeled training images (training set 1) have
been used to train the R-CNN network. In the second iteration, transfer learning is done based on the trained
network obtained from iteration 1 (itr1_net). Firstly, itr1_net has been used to perform car plate localization
on another set of training dataset (training set 2), which is made up of 232 training images. Then, based
on the localization results on training set 2, all the network’s false positive locations are labeled on training
set 2 as background class while the true car plate locations have been labeled as car plate class. The labeled
training set 2 is then used to train the R-CNN network in this second iteration of transfer learning. By doing
so, it reduces the false positive detection of the network. Lastly, further training is carried out based

Bulletin of Electr Eng & Inf ISSN: 2302-9285 
Malaysian car plate localization using region...(T. E. Liang)
413
on the trained network obtained from iteration 2 (itr2_net). In this iteration, 300 training images
from the training set 1 are relabeled so that both the car plate edges and the car plate characters are enclosed
by the ground-truth bounding boxes. The purpose of doing so is to allow the network to learn that the car
plate numbers are actually located on car plates. The relabeled training images are used to perform transfer
learning on the itr2 net.
(a) (b) (c)
Figure 1. (a-c) Examplesof labeled ground truth images used during training in iterations 1, 2, 3 respectively
2.3. Post-processing
The post-processing stage is required to perform filtering on the generated bounding boxes
by the R-CNN network during car plate localization testing phase. Firstly, confidence filtering stage
is applied to the generated bounding boxes. The R-CNN network produces bounding boxes as well as
the localization confidence factor for each of the classified car-plate region. The generated bounding boxes
is filtered with a preset confidence threshold t, where only the bounding boxes with a confidence factor
greater than the threshold value will be taken as the localization results. The threshold that has been chosen
is 0.92, where it has been obtained experimentally by executing car plate localization on training set 2.
If there are cases where none of the generated bounding boxes having a confidence factor greater than 0.92,
the next highest confidence bounding box is selected. By applying this filtering stage, the false positive
regions detected by the R-CNN network can be effectively eliminated. After tha t, all overlapping bounding
boxes, where any bounding boxes with intersection over union (IoU) ratio greater than threshold r
are merged. The threshold r that has been chosen is 0.15. As for the localization confidence factor
of the resultant merged bounding box, it is calculated as the average confidence of all the overlapping
bounding boxes.
2.4. Evaluation method
In order to evaluate the performance of the proposed system,several evaluation criteria are applied,
namely the number of true positives (TP), false positives (FP), false negatives (FN), precision rate and recall
rate [14]. These criteria have been used to quantify the accuracy of the car plate localization results
for the proposed system. TP is referred to as the detected regions that contain real car plates. In this paper,
the IoU ratio between the detected regions and the carplate ground-truth regions have been computed during
the evaluation stage. The detected regions are considered as TP if the IoU ratio between the detected regions
and the ground-truth is greater than 0.4. FP refers to the detected regions which do not contain any car plate.
FP shows the number of regions that have been incorrectly detected. FN refers to the miss rate i.e. real car
platesthat have not been detected by the system. We define the precision rate as the totalnumberof correctly
detected car plates over the total number of detected regions. It has been used to provide an insight into
the number of false alarms. A system with low precision rate will generate many bounding boxes with only
a few of them containing correctly detected car plates. We also define recall as the total number of correctly
detected car plates over the total number of ground-truth. This criterion gives information on the amount
of ground-truth car plates that have been successfully detected. If the system has been unable to localize
the majority of the ground-truth car plates,it will have a low recall rate.
2.5. Creating a benchmark
Two different approaches have been used to train another 2 R-CNN networks for car plate
localization. The car plate localization results produced by the 2 R-CNN networks are then being used as
a benchmark to compare with the results produced by AlexNet trained with transfer learning. In the first
approach, R-CNN network has been trained from scratch by using the captured car plate images. In this
approach, a 5 layer CNN model has been built, which consists of 3 convolutional layers and 2 fully
connected layers. Similar to the transfer learning approach, the last fully connected layer is made up of two
neurons in order to output 2 classes, which are the car plate class and also the background class. In this
approach, the R-CNN network has been trained with a total of 693 training images, which is about 60%
of the totalcaptured images. In the second approach,transferlearning based on CIFAR-10 network has been

 ISSN: 2302-9285
414
carried out. CIFAR-10 network is a pre-trained network that has been created by Alex Krizhevsky [24]. It is
a 5 layer CNN model that has been trained with 50.000 of training images. The training images are made up
of 10 classes, which are cat, dog, bird, horse, frog, deer, airplane, ship, automobile and truck. In this transfer
learning approach, the last few layers of the pre-trained network have been replaced with the layers that suit
for car plate localization. The first fully connected layer has been changed to 256 neurons instead
of the original 64 neurons in order to allow the network to relate more learned features to the characteristics
in different classes. As for the last fully connected layer, it has been changed to 2 neurons instead
of 10 neurons, which translates to 2 output classes, car plate, and background. Similarly, in this approach,
the R-CNN network hasalso been trained with 693 training images.
3. RESULTS
There are a total of 3 training approaches being applied in this work, namely, training a new
network from scratch (approach 1), transferlearning based on CIFAR-10 network (approach 2),and tra nsfer
learning based on AlexNet (approach 3). Figure 2 shows the training graphs for all three approaches.
Mini-batch is the number of training images to be propagated through the network during each iteration.
One mini-batch is a subset of the total training images available. When all the mini-batches have completely
passed through the network, it is referred to as one epoch. For the new network that is trained from scratch,
the total training time is 4.2 hours. From the training graph in Figure 2(a), it has been observed that
the mini-batch loss of the network saturates at 0.45 while the mini-batch accuracy saturates at 80%.
This is because the amount of training images available in this project is considerably small and hence,
the network accuracy cannot improve furtherwith this numberof training images.
(a)
(b)
(c)
Figure 2. (a-c) Mini-batch loss and mini-batch accuracy traininggraphs in approaches 1,2, 3 respectively

415
As for the CIFAR-10 network, the training time is 0.52 hour. The training time for this approach
is shorter because the network is running on an image input layer of smaller size. As a result, fewer
multiply-accumulate operations have been performed and less amount of memory space was needed.
From the training graph in Figure 2(b), it can be seen that the average mini-batch loss for the network is at
around 0.15 while the average mini-batch accuracy is at around 95%, which outperforms the first approach.
Figure 2(c) shows the training graph for transfer learning based on AlextNet in iteration 3. A total of 5.9
hours was taken to train the R-CNN network in this iteration. The time taken to train this network
is relatively long as the AlexNet is a big neural network that contains more layers and the image input layer
size is relatively bigger. As a result, more multiply-accumulate operations are performed and a higher amount
of memory space was used. Therefore, the total training time is longer. As illustrated in the training graph,
the mini-batch loss approaches 0 while the mini-batch accuracy approaches 100%, which reflects that
the trained network is able to achieve high accuracy.
3.1. Car plate localizationresults
In this work, a total of 232 test images were used to test the developed car plate localization system
(around 20% of the total acquired images in this work). Figure 3 shows the AlexNet car plate localization
results for car plates captured in different situations. The proposed approach is able to localize car-plates
in different situations accurately, e.g. even for incomplete car-plate (missing 1 car-plate character), special
Malaysian car-plate, multiple car-plates, occlusion and at an angle, as shown in Figure 3. Table 1 shows
the results of the evaluation criteria set. The average precision and recall are calculated by averaging
the precision and recall results for all test images. The “total detected regions” is the sum of bounding boxes
produced by the car plate localization system. The “average IoU with TP” is the average amount
of intersection between the detected TP with the manual labeled ground-truths. As for the “requires merging”
criteria, it is the total number of bounding boxes that required merging to be performed. Lastly, the “average.
TP confidence” is calculated by averaging the confidence factors foreach TP boundingboxes.
Table 1. Evaluation criteria results
Proposed
network
CIFAR
10
AlexNet
Iter. 1 Iter. 3
Average Precision (%) 19.4 82.22 93.64 95.19
Average Recall (%) 18.32 86.28 97.92 97.84
Total detected regions 232 295 290 289
Total true positives
(TP)
45 222 258 265
Total false positives
(FP)
187 73 32 24
Total false negatives
(FN)
226 54 14 11
Average IoU with TP 0.6061 0.6588 0.7525 0.7346
Requires merging 0 83 32 5
Average TP
confidence
0.5648 0.9856 0.9950 0.9964
Figure 3. Carplate localization results
Based on the results, it shows that the average precision and recall rate had increased from training
new network from scratch (approach 1) to three-iteration transfer learning based on AlexNet (approach 3).
The same trend hasbeen observed in the totalnumberof TP. This indicates that the system performance ha d
improved when moving from approach 1 to approach 3. On the other hand, the amount of FP and FN had
decreased from approach 1 to approach 3.This has shown the system has been able to produce better qua lity
results with lesser noise. Besides, the average IoU between TP and ground-truths have been recorded to be
at around 0.6 to 0.75, with the highest value achieved in approach 3. This has shown that the developed
system in approach 3 hasbeen able to produce boundingboxes that closely matched the labeled ground-truth.
From the evaluation criteria results, it can be observed that the highest precision and recall rates have been
achieved in approach 3,which are 95.19% and 97.84%.
The same trend is observed in the totalnumberof TP. Out of the 289 detected regions in approach 3,
there are 265 regions are true positives, which is 91.7% out of the total detected regions. This shows that
the proposed method is able to localize car-plates with low false detections. The average IoU between TP
and ground-truths have been recorded to be at 0.7346. In addition,it has been observed that the totalnumber
of bounding boxes that required merging in approach 3 is very low. This means that the R-CNN network

 ISSN: 2302-9285
416
is able to produce bounding boxes that completely encloses the car plate regions and less likely to require
the merging stage. As for the average TP confidence, the trained network is been able to achieve very high
confidence valuesfor the detected TP, close to 1.
3.2. Analysis
Based on the evaluation results, it can be seen that the R-CNN network has produced accurate car
plate localization results. Further training the AlexNet in iteration 3 was able to produce even better results as
compared to the first iteration. This is because the network in the third iteration has now learned that
car-plate numbers are found on car plates instead of just learning to recognize characters and numbers.
As a result, this has assisted the car plate localization system to identify the region of car plates within
the input images. The distribution graph of TP confidence factors for AlexNet in iteration 3 is shown in
Figure 4 further proves the localization accuracy of the network. From the distribution results, it can b e seen
that most of the confidence factors for the TP are greater than 0.99. It reflects that the iteration 3 network
is able to produce accurate localization results with high confidence factor on most of the test images.
When comparing the AlextNet result to CIFAR-10 network result, it can be observed that the CIFAR-10
network has relatively lower precision and recall rate. This is because AlexNet is a huge neural network that
has deep layers and has been trained with a very large amount of training dataset (1.2 million images) in
the previous training. As a result, AlexNet has a higher number of features being learned from the previous
training ascompared to CIFAR-10 network, which is being trained with 50.000 images.
On the other hand, for the newly built network that has been trained from scratch, weaker results
have been produced, where it has lower precision and recall rate. This is because the amount of training
dataset available fortraining the R-CNN network is considerably small as compared to the amount of training
dataset that the pre-trained networks have been trained on. As a result, without performing transfer learning,
purely training the new network from scratch with the small amount of dataset has resulted in lower car pla te
localization accuracy. This shows that if the amount of training dataset is not sufficient, it is preferable
to perform transfer learning instead of training a new network from scratch. Even though the AlexNet
in iteration 3 has been able to produce accurate car plate localization results, there are still some false
positives that have been detected as shown in Figure 5. Generally, the false positives detected by the system
can be divided into two main types,which are the detection of car model names and the background regions.
Some of the car model names are incorrectly detected as the car plate region as the car model name regions
also contained characters and numbers, which are very similar to the car plate regions. On the other ha nd,
background regions such as car grilles have also been incorrectly detected as car plate regions.
This is because the regions of the car grilles contained a well-defined boundary, which is identical
to carplates.
Figure 4. Distribution graph of TP confidence factorfor
iteration 3 network
Figure 5. Examples forfalse
positives
3.3. Comparison with existing work
In order to benchmark the developed car plate localization system, comparison with previous work
has been performed. Currently, there is no deep learning based work that has been done on Malaysian
car-plates. Besides, none of the previous work is using R-CNN approach to perform car plate localization.

417
Therefore, a comparison has been done against the previous work that is not exactly the same but closest
to the current work. Even though there are some limitations on this comparison as it is not one to one
comparison, but it served as an indicator to provide a general view on the current standing of this work.
The previous work done by Hui Li and Chunhua Shen [14] has been chosen as the benchmark work
for comparison. Table 2 illustrates the comparison between [14] with the current work. The authors
in the previous work have carried out an evaluation on their system by running a car plate localization
process on Caltech dataset [25].
The Caltech dataset is made up of 126 US car plate images. All the images in the dataset are
the images of cars captured from the rear and acquired in the parking lots of California Institute
of Technology (Caltech). In order to compare the current work with the previous work, the proposed work
was also tested on the Caltech dataset. Based on the precision and recall rates as shown in Table 2, even
though the proposed method was purely trained based on Malaysian car plates, the system is still able to
produce comparable results to the previous work when applied on the Caltech dataset. This shows that
the R-CNN in the current work is sufficiently generalized to be able to work on different datasets. As for
execution performance, the processing time for the previous work in [9] (NVIDIA Tesla K40c, 6GB)
is 5 sec, while the proposed work’s processing times are 18 sec (NVIDIA GeForce 940MX, 4GB) and 4.5 sec
(NVIDIA GeForce GTX1060, 6GB) respectively. Re-running the proposed work on NVIDIA GeForce
GTX1060 is able to achieve similar processing time as compared to previous work. This has shown that
the deep learning algorithm processing time is heavily dependent on the hardware resources available.
Figure 6 shows some examples of localization results for the proposed method on the Caltech dataset. It can
be observed that the proposed method has been able to accurately localize car-plates from the input images
even though the images are having a US plate format.
Table 2. Comparison with previous work
Readingcar License Plates usingDeep Convolutional
Neural Networks andLTSM(H. Li andC. Shen,2016)
Current Project
Hardware NVIDIA Tesla K40cwith 6GB memory NVIDIA GeForce940MXwith 4GB memory
Network
architecture
Cascaded CNN with false positive eliminationby
handcraftedfeatures
a. CNN1:
− 4-layers 37-class CNN fortext detection
b. Refine bbox:
− geometric constraints + vertical edge detection
usingSobel operator
c. CNN2:
− 4-layers 2-class CNN for plate/non-plate detection
RCNN with 3 trainingiterations
a. TransferlearningfromAlexnet
b. Region proposal network:
− Propose2000regions that most probably
contain object
c. CNN:
− 8-layers 2-class Alexnet CNN forplate
detection
Training
dataset
CNN1:
a. 1.38 × 105
character images + 9 × 105
non-character
images
CNN2:
a. 3000 car plates fromdifferent countries
b. 5 × 104
synthesizedlicenseplates
c. 4 × 105
background images
All trainingimages are Malaysiancar plates
Trainingiteration1:
a. 689 trainingimages withlabelledgroundtruth
(trainingset 1)
Trainingiteration 2:
& background(trainingset 2)
Trainingiteration3:
that includes plate edges (from trainingset 1)
Evaluationon
Caltech
dataset (USA)
Precision : 97.56
Recall : 95.24
Precision : 84.26
Recall : 94.44
Processing
time
5 sec
18 sec
4.5 sec (rerunon NVIDIAGeForce GTX1060
6GB)
Figure 6. Carplate localization results on the Caltech dataset

 ISSN: 2302-9285
418
4. CONCLUSION
In this work, a Malaysian car plate localization system has been successfully developed using deep
learning approach. The developed car plate localization system has been able to accurately localize the car
plate location from input images by creating bounding boxes around the car plates. The images used in this
work include images with multiple car plates, special character car plates, car plates with logos on it, partially
blocked car plates and car plates captured at different angles. There are several potential improvements
and modificationsthat can be done to further enhance the current carplate localization system. Firstly, a ca r
plate number recognition stage can be added to the end of the car plate localization stage. By adding the car
plate number recognition stage, the car plate numbers in the localized regions can be recognized
and at the same time, any “background” type of false positives produced by the localization stage can be
eliminated effectively. In addition, a large number of car model name images can be collected and be used
to re-train the R-CNN network as a background class. This can effectively help the network to reject those
text-based false positive regions especially the regions containing car model name. By doing so, the system
can be made more robust as it will now be able to work in different weatherand lighting conditions.
ACKNOWLEDGEMENTS
The authors thank the Ministry of Education Malaysia and Universiti Tun Hussein Onn (UTHM)
for their support under the Fundamental Research Grant Scheme (FRGS), grant number 1580 and Universiti
Teknologi Malaysia underResearch University Grant (GUP), grant number19H61.
REFERENCES
[1] P. I. Reji and V. S. Dharun, “License Plate Localization: A Review,” International Journal of Engineering Trends
and Technology (IJETT), vol. 10, no. 13, pp. 604-615, Apr. 2014.
[2] T. K. Lai, A. F. Abbas, A. M. Abdu, U. U. Sheikh, M. Mokji, and K. Khalil, “Super Resolution of Car Plate
Images U sing Generative Adversarial Networks,” 2019 IEEE 15th International Colloquium on Signal Processing
& Its Applications (CSPA), Penang, Malaysia, pp. 80-85, 2019.
[3] J. Tai, S. Tseng, C. Lin, and K. Song, “Real-Time Image Tracking for Automatic Traffic Monitoring
and Enforcement Applications,” Image and Vision Computing, vol. 22, no. 6, pp. 485-501, June 2004.
[4] S. Setumin, U. U. Sheikh, and S. A. R. Abu-Bakar, “Character-Based Car Plate Detection and Localization,” 10th
International Conference on Information Science, Signal Processing and their Applications (ISSPA 2010),
Kuala Lumpur, pp. 737-740, 2010.
[5] S. Setumin, U. U. Sheikh, and S. Member, “Car Plate Character Extraction and Recognition using Stroke Analysis
Computer Vision, Video and Image Processing Lab., Dept.. of Microelectronics and Computer,” 2010 Sixth Int.
Conf. Signal-Image Technol. Internet Based Syst., pp. 30-34, 2010.
[6] V. Ganapathy and W. L. D. Lui, “A Malaysian Vehicle License Plate Localization and Recognition System,”
Journal Systemics, Cybernetics and Informatics, vol. 6, no. 1, pp. 13-20, 2008.
[7] P. Tarabek, “A Real-Time License Plate Localization Method Based on Vertical Edge Analysis,” 2012 Federated
Conference on Computer Science and Information Systems (FedCSIS), Wroclaw, pp. 149-154, 2012.
[8] S. Saha and S. Basu, “License Plate Localization from Vehicle Images: An Edge Based Multi-Stage Approach,”
International Journal of Recent Trends in Engineering, vol. 1, no. 1, pp. 284-288, May 2009.
[9] M. H. Dashtban, Z. Dashtan, and H. Bevrani, “A Novel Approach for Vehicle License Plate Localization
and Recognition,” in International Journal of Computer Applications, vol. 26, no. 11, pp. 22-30, July 2011.
[10] H. Kwaśnicka and B. Wawrzyniak, “License Plate Localization and Recognition in Camera Pictures,” Artificial
Intelligence Methods, pp. 243-246, November 2002.
[11] B. V. Ghaffari, T. Kitajima, and M. Kouhnavard, “A New Localization Method for Malaysian License Plate,”
GSTF J. Comput., vol. 2, no. 4, 2013.
[12] L. Deng and D. Yu, “Deep Learning: Methods and Applications,” Foundations and Trends® in Signal Processing,
vol. 7, no. 3-4, pp. 197-387, 2013.
[13] Y. LeCun, Y. Bengio, G. Hinton, L. Y., B. Y., and H. G., “Deep Learning,” Nature, vol. 521, no. 7553,
pp. 436-444, May 2015.
[14] H. Li and C. Shen, “Reading Car License Plates using Deep Convolutional Neural Networks and LSTMs,” Arxiv,
pp. 1-17, Jan. 2016.
[15] S. A. Radzi and M. Khalil-Hani, “Character Recognition of License Plate Number using Convolutional Neural
Network,” Visual Informatics: Sustaining Research and Innovations (IVIC), pp. 45-55, January 2011.
[16] A. Naimi, Y. Kessentini, and M. Hammami, “Multi-nation and Multi-norm License Plates Detection in Real
Traffic Surveillance Environment using Deep Learning,” International Conference on Neural Information
Processing Neural Information Processing, vol. 9948, pp. 462-469, September 2016.
[17] M. Earl, “Number plate recognition with Tensorflow.” [Online]. Available:
https://matthewearl.github.io/2016/05/06/cnn-anpr/. [Accessed: 16-Nov-2016].
[18] C. Gerber and M. Chung, “Number Plate Detection with a Multi-Convolutional Neural Network Approach with
Optical Character,” in Journal of Information Processing Systems, vol. 12, no. 1, pp. 100-108, March 2016.

419
[19] The Mathworks, “Image Processing ToolboxTM User’s Guide R2016b,” 2016.
[20] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural
Networks,” in Advances in neural information processing systems, vol. 25, no. 2, pp. 1-9, Januay 2012.
[21] R. Girshick, J. Donahue, T. Darrell and J. Malik, "Region-Based Convolutional Networks for Accurate Object
Detection and Segmentation," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 38, no. 1,
pp. 142-158, 1 Jan. 2016.
[22] J. Carreira and C. Sminchisescu, "CPMC: Automatic Object Segmentation Using Constrained Parametric
Min-Cuts," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, no. 7, pp. 1312-1328,
July 2012.
[23] J. Uijlings, K. van de Sande, T. Gevers, and A. Smeulders, “Selective Search for Object Recognition,”
International Journal of Computer Vision, vol. 104, no. 2, pp. 154-171, April 2013.
[24] A. Krizhevsky, “Learning Multiple Layers of Features from Tiny Images,” Sci. Dep. Univ. Toronto, pp. 1-60,
April 2009.
[25] C. Computational Vision Group, “Computational Vision: Archive,” 1999. [Online]. Available:
http://www.vision.caltech.edu/html-files/archive.html. [Accessed: 28-May-2017].
BIOGRAPHIES OF AUTHORS
Tay Eng Liang received the Bachelor of Electrical and Electronics Engineering (Hons.) from
University Tenaga Nasional, Malaysia, in 2015, and Master of Engineering in Computer and
Microelectronic Systems from Universiti Teknologi Malaysia, Malaysia, in 2017. He was
the recipient of Industry Excellence Award of University Tenaga Nasional in 2015 and Best
Student Award from Universiti Teknologi Malaysia in 2017. In 2015, he has joined Intel
Microelectronics (M) Sdn. Bhd, Penang, Malaysia, as a design engineer. He has been working
on silicon design and involved in designing the next generation desktop processor. He is
focusing in static timing analysis domain and is leading the timing convergence activity for
the design.
Usman Ullah Sheikh received his PhD degree (2009) in image processing and computer vision
from Universiti Teknologi Malaysia. His research work is mainly on computer vision, machine
learning and embedded systems design. He is currently a Senior Lecturer at Universiti Teknologi
Malaysia.
Mohd Norzali Haji Mohd is currently working as Senior Lecturer, Faculty Lab Manager (HoD)
and Former Industrial Training Coordinator at the Department of Computer Engineering, Faculty
of Electrical and Electronic Engineering, Universiti Tun Hussein Onn Malaysia (UTHM). He
received B.Eng., M.Eng. from Fukui University, Japan in 2002 and 2004 respectively. In April
2015, he received Ph.D. from the Department of Information Sciences and Biomedical
Engineering, Kagoshima University, Japan. His research interests are computer vision, robot
vision, pattern recognition and image processing.

Malaysian car plate localization using region-based convolutional neural network

More Related Content

What's hot (20)

Similar to Malaysian car plate localization using region-based convolutional neural network (20)

More from journalBEEI (20)

Recently uploaded (20)

Malaysian car plate localization using region-based convolutional neural network