License Plate Localization

Using Phase Stretch Transform

Waqas Ahmed Shehzad Khalid Imran Ahmed Siddiqi
Department of Computer Engineering Department of Computer Engineering Department of Computer Science
Bahria University Bahria University Bahria University
Islamabad, Pakistan Islamabad, Pakistan Islamabad, Pakistan
waqas qasi2000@yahoo.com shehzad khalid@hotmail.com imran.siddiqi@gmail.com

Abstract—Intelligent Transport System (ITS) is an important

part of today’s modern society. With increasing number of
vehicles and road networks impose the requirement of highly
refined and precise Intelligent Transport System. Automatic Fig. 1. Automatic License Plate Recognition (ALPR)
License Plate Recognition (ALPR) is also included in the main
components of ITS, which uses image processing techniques
to detect and recognize vehicles on the basis of license plates. After receiving the captured images ALPR software pro-
License plate localization is the most complex and important
cessed for image enhancement in preprocessing stage and
stage in license plate recognition due to variation of license
plate(LP) designs, font styles and illumination conditions. Since, then license plate detection/localization will be done to ac-
all the subsequent stages of LP recognition greatly depend on curately locate the LP regions in captured images. After
the accurate detection of license plates. A novel approach is LP detection stage, identified plate regions are segmented to
presented in this paper for the localization of license plates with extract characters which are then recognized as string by some
various sizes in different illumination conditions and complex
artificial intelligence techniques without human intervention.
backgrounds. Phase Stretch Transform (PST) is utilized for edge
enhancement and detection. Then license plates candidate regions The extracted license plate number in form of string is stored
are generated by applying morphological operations. Finally, the in database along with image for further processing and review
license plate is filtered out and validated on the basis geometrical depending on the use of ALPR application.
properties of candidate regions. The performance of proposed In this paper, we focus on localization of license plate in still
methodology is tested on NTUA Media Lab LPR dataset and
images with complex backgrounds and different illumination
achieves Precision of 89.98% , Recall of 93.71% and detection
rate of 98.86%. The experimental results of proposed approach conditions. We propose a novel approach that utilizes phase
demonstrate the effectiveness of PST based license plate detection congruency based method for edge detection in LP images.
in captured images. This method employ Phase Stretch Transform (PST) to locate
Index Terms—License Plate Recognition, License Plate Local- license plate edges even in low contrast image with complex
ization, Phase Stretch Transform
background and varying lighting conditions. We evaluated this
approach on a NTUA LPR dataset because vehicle images are
I. I NTRODUCTION
categorized into 7 classes depending on various conditions that
Automatic License Plate Recognition (ALPR) systems allow are required to prove the effectiveness of proposed approach.
to locate and automatically extract vehicle license plates for The outcomes of proposed method are comparable with other
recognition. These systems are extensively used for wide hand-engineered techniques presented in literature.
range of commercial applications such as traffic monitoring, The rest of paper include the following sections. Section II
service theft monitoring, toll collection, law enforcement and provides a brief overview on related work. Methodology that
control access to parking facilities. Vehicle speed and lane is proposed for license plate detection using Phase Stretch
violations mostly monitored by applying ALPR system for Transform(PST) is explained in section III, followed by the
law enforcement [1][2] other includes travel time monitoring discussion on experimental results in section IV. Finally,
[3]. The main advantage of such kind of system for law section V presents the conclusion of our study.
enforcement is that it can store image record for reference
use and as a proof of violation. Self-service petrol stations II. R ELATED W ORK
successfully utilizes ALPR system for service theft monitoring Many license plate localization algorithm along with seg-
by tracking registration of vehicles using license plate number. mentation and recognition techniques have been proposed in
ALPR system take picture of front and back license plates literature for the implementation of ALPR system. A brief
by installed cameras. The captured images of vehicle is read overview about traditional LP detection approaches is pre-
by frame grabber which sends pixel information to ALPR sented in this section. LP detection algorithms can be broadly
software for further analysis. The process for license plate categorized into four approaches: edge-based, color-based,
recognition is depicted as texture-based and machine learning(ML) based approaches.
 Most commonly edge based license plate detection tech- is used for license plate localization in [15].
niques are used in literature. In [4], edge information is In recent times, machine learning based techniques are
utilized for plate detection. The horizontal and vertical edges applied for the detection of license plate. A large number of
are extracted by applying edge detection operators on the classification based methods such as artificial neural networks
image after image smoothing and enhancement operations. (ANNs) [20], genetic algorithm (GA) [21][22], and support
Then rectangular areas which contains high edge density are vector machine (SVM) [23] were mostly implemented for
makred as candidiate regions for license plate. Subsequently, license plate detection. All these techniques requires proper
rule based mechanism is followed using threshold to left training and parameter tuning to produces decent classification
many candidate regions. At end, final license plate region is results. In [24], Adaptive boosting (AdaBoost) is employed
selected by applying Connected Component Analysis (CCA) with Haar-like features in cascaded framework for the locating
on candidate regions. In [5] [6], same mechanism is followed license plate region. The results of AdaBoost based technique
to detect license plate region. In [7], license plates are detected are robust even in presence of variaton in illumination con-
by applying novel line density filter on still images. In this ditions, color, size, and position of license plate due to the
method, regions with high edge density are connected together use of Haar-like features. HOG based features with AdaBoost
while sparse regions are removed by applying line den- are also utilized in [25] for the locating license plate position
sity filter in horizontal(row) and vertical(column) directions. in capture images. Recently, deep learning based methods are
Then license plates candidates are verified by using color also employed for the localization of vehicle number plates
saliency features with linear SVM. Also techniques presented in images. In [26], convolutional neural network (CNN) based
in [9][10][11] utilizes edge statistic based scheme for the scheme is introduced number plate detection. In this work,
localization of license plate region. Similarly, edge magnitude two detectors are trained by using cropped characters from
and variance can be calculated for image divided into blocks. simple text and cropped license plates from vehicle images for
The blocks containing high value for these statistical measures character based and license plate based detection. Whereas, a
are declared as the candidate region for license plate [12]. novel approach based on Fast-YOLO and YOLOv2 for the
The edge detection techniques along with morphological op- detection cars and LPs in real world environment is presented
erations yields better estimates of plate regions. However, all in [27]. Also, region based CNN are employed in literature
such techniques give better results only when edges produced for the detection of license plates[28].
by license plate characters are clear. Due to presence of any
inclination or affine transformation or image blur in license III. P ROPOSED M ETHOD
plate frame these algorithms gives poor results. Ability to correctly localize the license plate region is of
In some countries, color of text characters, logos and prime importance in LPR system. The subsequent stages of
background are fixed. Thus, color information from the image segmentation and recognition are dependent on accuracy LP
is used to detect license plate region. Approaches presented in detection process. In general, License plates consists of several
[16][17] uses color information for the locating license plate in characters and digits having high contrast with background.
images. However, these approaches give unpredictable results Which results in high density of edges and texture variation
due to varying lighting condition which may disrupt the color in LP region. Therefore, edge information is mostly utilized
information. A novel approach Sliding Concentric Windows in localization of license plate region. However, most of the
(SCW) was intoduced by Deb et al. in [18], which utilizes license plate localization approaches build upon edge detection
three step procedure for the detection of license plates. In shows poor performance in case of low contrast images. In this
first step, sliding concentric windows are employed for the paper, Phase Stretch Transform based algorithm is utilized for
detection of possible candidate regions. Then, HSI color based the detection of edges. This method is relatively stable and
verification scheme is applied on candidate regions. Finally, robust in presence of different license plate sizes, illumination
position histogram are used to segment out license plate conditions and complex background. Figure 2 presents the
characters. In [19], HSV color space based image is partitioned architecture of our PST based LP detection method. The
into smaller regions and then each region is inspected for proposed approach is composed of five major stages.
license plate regions by using filtering procedure. • Preprocessing of captured image;
Also, methods based on Hough transform, wavelet trans- • Phase Stretch Transform for edge detection;
form and Gabor filters are also employed for LP detection. • Mask generation for elimination of undesired regions;
As Hough transform is mostly applied in image processing • Morphological operations for the generation of LP can-
to detect straight lines. Therefore it can be easily utilized for didates; and
the detection of rectangular shape license plates [13]. Gabor • Validation of number plate region.
filters based approach is presented in [14] for license plate
region detection in captured images with fixed angle. This A. Image Pre-Processing
approach analyze texture with different scale and direction. The proposed license plate localization (LPL) system re-
However, transformation based license plate detection methods ceive the input image captured by camera and passes it for pre-
are computationally expensive and requires image with simple processing. This stage involves conversion of color image into
background for high detection rates. Vector quantization (VQ) grey-level image and smoothing of image for noise reduction.
 In eq.(1), Frequency response of localization kernel and
non-linear frequency dependent phase kernel is denoted by
L[u, v] and K[u, v] respectively [29]. The variables m and n
are two dimensional spatial variable whereas u and v are the
frequency variables. L[u, v] is set to 1 for simplification and
K[u, v] is given by frequency dependent phase as follows:
K[u, v] = ej.ϕ[u.v] (4)

ϕ[u, v] = ϕP olar [r] (5)

and
W.r. tan−1 W.r − (1/2). ln (1 + (W.r)2 )
ϕP olar [r] = S.
W.rmax . tan−1 W.rmax − (1/2). ln (1 + (W.r)2 )
√ (6)
Where r = u2 + v 2 , θ = tan−1 (v/u) and rmax is the
maximum value of r. The variable W and S corresponds to
the wrap and strength of the phase profile.
In this work, spatial kernel is applied first to cater the noise
problem then PST kernel is utlized for phase enhancement and
Fig. 2. Flow diagram for License Plate Detection edge detection. Based on empirical evaluation, PST parameters
phase strength (S) and warp strength (W) are set to 0.7
and 25 respectively. Finally, post processing operations using
First, input vehicle image is converted into gray scale image thresholding and series of morphological steps are performed
which is represented by the single value of brightness in image. to generate PST image[29]. The output PST image contain
Then for image enhancement, Gaussian smoothing filter is enhanced edges of license plate region as well as other sharp
used to eliminate adaptive noise present in the image. intensity regions as shown in Figure 3 (d), whereas Figure
3 (c) illustrates the binary image without applying PST edge
B. Phase Stretch Transform
enhancement.
The phase stretch transform or PST is a physic inspired
technique that is used for feature detection and classification C. Mask generation
in images. PST is derived from the work on time stretch dis- At this stage, mask is generated to remove unwanted regions
persive Fourier transform, which implements two-dimensional from PST image to accurately locate the license plate. In
phase function in frequency domain. The frequency content this regard, Adaptive Binarization is performed on smoothed
of image determine the amount of phase applied by phase
function. Small amount of phase is applied to low frequency
region in the image while higher frequency components get
the higher phase. Therefore, large amount of phase is applied
to edges as they possess higher frequency contents. The Phase
Stretch Transform(PST) for 2D image B[m,n] is defined as

D E
˜ {K[u, v].L[u, v].F̂ {B[m, n]}}
A[m, n] = ] IF (1)

Where A[m,n] is the output phase image, ] represents the

angel operation. F̂ {.} is Fourier transform and IF˜ {.} is
inverse Fourier transform for two dimensional images.

M −1 N −1
1 X X
B[m, n]e−j2π( M m+ N n)
u v
F̂ {B[m, n]} = F [u, v] =
M N m=0 n=0
(2)

N −1 N −1
F [u, v]ej2π( M m+ N n)
X X u v
˜ {F [u, v]} = B[m, n] =
IF
m=0 n=0
(3) Fig. 3. Edge enhancement using Phase Stretch Transform (PST)
 Fig. 4. License Plate Detection Process

image acquired after the first stage of image preprocessing. results in an image containing candidate license plate regions.
As fixed global threshold is not suitable for binarization of the
image because of spatial, temporal and illumination variations
in the captured image. Therefore, a local adaptive threshold E. Validation of plate region
is computed dynamically by averaging the pixel intensities In this step, connected component analysis (CCA) is em-
of neighborhood pixels at each pixel position in an image as ployed using 8-connected neighborhood to identify and consti-
described in [30]. A simple mechanism is utilized to allocate tute the set of regions with main geometrical properties such
pixels to foreground or background of an image. In this as height, width, area, aspect ratio, extent, eccentricity and
respect, if the pixel gray value exceeds the 90% of local orientation. Further in depth details regarding the geometric
average then it is allocated to foreground; otherwise, it is properties of license plates could be found in [31]. After that,
marked as background pixel. We then perform morphological candidate regions are filtered on the basis of computed prop-
closing operation followed by complement to obtain the binary erties using prior knowledge of License Plates as threshold.
mask. At that point, generated mask using AND operation This filtering based on geometrical properties significantly
is applied on edged enhanced PST image to eliminate the decreases the number of candidate regions. However, there
unwanted details as illustrated in Figure 4(d). may be some false positives which do not represents the actual
LP region. For this reason, some validation criteria must be
D. Morphological operations for the selection of candidate incorporated to remove these false regions. In this approach,
regions connected component labeling is applied for validation on
During the previous step most of the useless information is each candidate license plate region. The number of connected
eliminated from the image. We further employ morphological component for each candidate is counted and only those
closing operation to connect the pixels in plate region. In this components that agrees the geometrical properties such as
closing operation structuring element is selected by using the height, width and aspect ratio of the character on the license
prior knowledge of maximum height and width of characters plate, are considered as valid LP components. Finally, false
in license plates. Also, this closing operation may connect plate regions are eliminated and required number plate region
the nearby regions or pixels with possible plate region. For is located in image by using acquired LP region coordinates.
this reason, morphological closing is followed by opening In case when no license plate is found in acquired image,
operation to separate apart the weak connected regions, which there may be the possibility that valid mask could not be
 TABLE I
P ERFORMANCE OF PROPOSED L ICENSE P LATE D ETECTION S YSTEM

End to End LP Recognition [31] Proposed Approach

Category Total Images Precision Recall F-Measure Precision Recall F-Measure
Color 67 82.54 92.50 87.24 88.38 97.09 92.53
Gray Scale 50 93.26 91.09 92.16 92.14 94.67 95.84
Color Large 138 90.11 94.01 92.02 91.38 95.84 93.55
Day Blurred 7 65.30 85.40 74.01 62.59 67.82 65.10
Shadows 52 85.41 92.31 88.73 86.24 91.70 88.88
Day Close View 122 76.77 89.12 82.49 91.34 91.43 91.38
Night 3 90.67 91.01 90.84 99.75 91.65 92.52
Total 439 87.68 91.30 89.45 89.98 93.71 91.81

TABLE II The precision is the ratio of total number of pixels which

C OMPARISON OF L ICENSE P LATE D ETECTION R ATES are correctly labeled as license plate pixel to total number of
Experiment Number Correct Missed Detection pixel labeled as license plate pixels by proposed method.
of Test Detection Detection Rate
Samples
Proposed (GT ∩ DR)
335 334 1 99.70 P recision = (7)
Approach DR
Peter
Tarabek 335 333 2 99.40 Recall or true positive rate is the ratio of total number of
[32] pixels which are correctly labeled as license plate pixel to total
G. Abo number of pixel labeled as license plate in ground truth.
Samra et. al 335 327 8 97.61
[21]
(GT ∩ DR)
Recall = (8)
GT
generated due to non uniform illumination condition in the
image . Therefore, algorithm repeat the steps of morphological F-measure combines precision and recall into one metric
operation with LP candidate selection and validation on PST using harmonic mean.
based enhanced image without masking operation to locate
desired number plate. (P recision × Recall)
F M easure = 2× (9)
(P recision + Recall)
IV. E XPERIMENTAL R ESULTS AND A NALYSIS
This section contain the details of dataset utilized for Whereas, Success Rate/Detection Rate is computed as the
the testing of proposed framework with baseline metrics to percentage of successful detected license plates out of total
compute and compare effectiveness of proposed technique. presented plates for detection.
Also, here we present comparison of proposed method with C. Performance Evaluation
already proposed techniques in literature.
In order to evaluate the effectiveness of presented approach,
A. Dataset all images of MediaLab LPR dataset are utilized and resized
The proposed framework is tested using LPR database into 800 pixel in width while maintaining the aspect ratio to
provided by MediaLab of National Technical University of ensure higher processing speed for the detection of license
Athens (NTUA) [33]. These images contain vehicles with plates. The proposed method achieves better results in terms
Greek license plates in different scenes with various illu- of the Precision, Recall and Detection Rate, where most of
mination conditions from front and rear view. The number license plates regions were identified effectively. Two exper-
plates in captured images starts with 3 alphabets, hyphen iments were performed. In first experiment we utilized the
character and then followed by 4 digits in a single line over whole dataset of 439 images, while in second experiment,
white background. The dataset consist of 439 images having only 335 images of four categories (color images large, LP
7 different categories as illustrated in Table 1. with shadows, Day close view, and Shadows in plate) are
used. Table 1 and 2 presents result of proposed method in
B. Performance Metrics experiment 1 & 2 respectively, in terms of computed statistical
The performance metrics used for the analysis and compari- measures such as Precision, Recall, F-Measure and Detection
son of proposed approach are Precision, Recall, F Measure and Rate for MediaLab LPR dataset. It is observed that proposed
Success Rate/Detection Rate. The area of the detected region technique attains higher value of 89.98% for Precision, value
(DR), ground truth region (GT), and the region obtained by of 93.98% for Recall and value of 91.94% for F-Measure.
intersection of GT and DR are used to compute performance These performance measures are computed on the basis of
statistics like Recall, Precision and F-metric for the proposed overlapping areas between predicted region and ground truth,
method, which are as follows. therefore do not produces 100% results because of slight over
 V. C ONCLUSION

We discussed the license plate detection using PST in detail

in presence of different image condition such as complex
background, different license plate sizes and illumination con-
dition, in this paper. The proposed framework for automated
license plate detection is based on edge enhancement and
detection using Phase Stretch Transform (PST). After edge
detection candidate regions are generated by using morpho-
logical operations. Then, these candidate regions are filter
out to locate license plate on the basis of geometric features
like length, width, and area etc. Also, the proposed tech-
nique provides accurate results even in low contrast images,
(bright) overexposed images and (dark) underexposed images.
However, the performance of proposed approach undergoes
by motion blur in the acquired image, which is the inherent
deficiency of techniques that are based on edge detection. It
is also confirmed during experiments that proposed method
can effectively detect and localize license plates of different
sizes from various angles and positions. The performance of
proposed approach shows the effectiveness of Phase Stretch
Transform (PST) based license plate detection for captured
images at various practical situations.

Fig. 5. Results of Proposed Technique in various conditions R EFERENCES

[1] D. C. Luvizon, B. T. Nassu, and M. R., “Vehicle speed estimation by

license plate detection and tracking,” in IEEE Int. Conf. Acoust., Speech,
Signal Process. (ICASSP), May 2014, pp. 6563–6567.
[2] H. Karwal and A. Girdhar, “Vehicle number plate detection system for
or under LP region detection, even the detection appears to indian vehicles,” in IEEE Int. Conf. Comput. Intell. Commun. Technol.
be correct at a galance. However, effectiveness of proposed (CICT), Feb. 2015, pp. 8–12.
[3] G. Liu, Z. Ma, Z. Du, and C. Wen, “The calculation method of road
approach is indicated by the achieved results. This is mainly travel time based on license plate recognition technology,” Advances in
due to the induction of Phase Stretch Transform for edge Information Technology and Education, pp. 385–389, 2011.
enhancement and detection in proposed method, which also [4] V. Karthikeyan and V. J. Vijayalakshmi, “Localization of license plate
using morphological operations,” arXiv preprint arXiv:1402.5623, 2014.
gives better results for edge detection in low contrast images as [5] H. Zhang, W. Jia, X. He, and Q. Wu, “Learning-based license plate
shown in Figure 5(j). In addition, the proposed work involves detection using global and local features,” in Pattern Recognition, 2006.
use of Gaussian smoothing filter which reduces computational ICPR 2006. 18th International Conference on, vol. 2. IEEE, 2006, pp.
1102–1105.
complexity and requires lesser time as compared to approach [6] P. V. Suryanarayana, S. K. Mitra, A. Banerjee, and A. K. Roy, “A mor-
presented in [31], which utilizes bilateral filter for image phology based approach for car license plate extraction,” in INDICON,
preprocessing. LP detection rate of proposed technique is 2005 Annual IEEE. IEEE, 2005, pp. 24–27.
99.70%, which is clearly higher than technique presented by [7] Y. Yuan, W. Zou, Y. Zhao, X. Wang, X. Hu, and N. Komodakis,
“A robust and efficient approach to license plate detection,” IEEE
P. Tarabek et al [32] and G. A. Samra et. al [21]. Whereas Transactions on Image Processing, vol. 26, no. 3, pp. 1102–1114, 2017.
in case of 439 images detection rate is decreased to 98.86%, [8] P. Comelli, P. Ferragina, M. N. Granieri, and F. Stabile, “Optical
due to presence of extremly blurred images. Also, proposed recognition of motor vehicle license plates,” IEEE transactions on
Vehicular Technology, vol. 44, no. 4, pp. 790–799, 1995.
approach do not have any training phase whereas P. Tarabek [9] J. Cano and J. Pérez-Cortés, “Vehicle license plate segmentation in
et al [32] uses 89 random samples to train the system, which natural images,” Pattern Recognition and Image Analysis, pp. 142–149,
are again included in testing phase. Figure 5 depicts some of 2003.
[10] J. Barroso, E. Dagless, A. Rafael, and J. Bulas-Cruz, “Number plate
examples from the detected license plates. reading using computer vision,” in Industrial Electronics, 1997. ISIE’97.,
Proceedings of the IEEE International Symposium on. IEEE, 1997, pp.
The blurred images has worse performance due to difficulty 761–766.
in extracting edges, whereas large images exhibits better per- [11] S. Shaikh, B. Lahiri, G. Bhatt, and N. Raja, “A novel approach for
automatic number plate recognition,” in Intelligent Systems and Signal
formance because of large resolution. On the other hand, night Processing (ISSP), 2013 International Conference on. IEEE, 2013, pp.
images are fairly easily to detect due to better illumination 375–380.
conditions caused by flashlight of camera. Also, problems [12] H. J. Lee, S. Y. Chen, and S. Z. Wang, “Extraction and recognition
of license plates of motorcycles and vehicles on highways,” in Pattern
arises in detection of LP in shadow images when shadows Recognition, 2004. ICPR 2004. Proceedings of the 17th International
are the part of license plates. Conference on, vol. 4. IEEE, 2004, pp. 356–359.
[13] S. Rasheed, A. Naeem, and O. Ishaq, “Automated number plate recog-
nition using hough lines and template matching,” in Proceedings of the
World Congress on Engineering and Computer Science, vol. 1, 2012,
pp. 24–26.
[14] F. Kahraman, B. Kurt, and M. Gökmen, “License plate character
segmentation based on the gabor transform and vector quantization,”
in ISCIS. Springer, 2003, pp. 381–388.
[15] R. Zunino and S. Rovetta, “Vector quantization for license-plate location
and image coding,” IEEE Transactions on Industrial Electronics, vol. 47,
no. 1, pp. 159–167, 2000.
[16] M. Ligang, W. Fengwen, and W. Han, “Automatic license plate detection
based on edge density and color model,” in Control and Decision
Conference, 2009. CCDC’09. Chinese. IEEE, 2009, pp. 3718–3721.
[17] M. U. M. Bhutta, H. Mahmood, and H. Malik, “An intelligent approach
for robust detection and recognition of multiple color and font styles
automobiles license plates: A feature-based algorithm,” in Audio, Lan-
guage and Image Processing (ICALIP), 2014 International Conference
on. IEEE, 2014, pp. 956–961.
[18] K. Deb and K. H. Jo, “Hsi color based vehicle license plate detection,”
in Control, Automation and Systems, 2008. ICCAS 2008. International
Conference on. IEEE, 2008, pp. 687–691.
[19] R. Azad, F. Davami, and B. Azad, “A novel and robust method for
automatic license plate recognition system based on pattern recognition,”
Adv. Comput. Sci., Int. J., vol. 2, no. 3, pp. 64–70, 2013.
[20] N. More and B. Tidke, “License plate identification using artificial
neural network and wavelet transformed feature selection,” in Pervasive
Computing (ICPC), 2015 International Conference on. IEEE, 2015,
pp. 1–5.
[21] G. A. Smara and F. Khalefah, “Localization of license plate number
using dynamic image processing techniques and genetic algorithms,”
IEEE transactions on evolutionary computation, vol. 18, no. 2, pp. 244–
257, 2014.
[22] J. Muhammad and H. Altun, “Improved license plate detection using
HOG-based features and genetic algorithm,” in 24th Signal Process.
Commun. Appl. Conf. (SIU), May 2016, pp. 1269–1272.
[23] W. Kusakunniran, K. Ngamaschariyakul, C. Chantaraviwat, K. Janvit-
tayanuchit, and K. Thongkanchorn, “A thai license plate localization
using svm,” in Computer Science and Engineering Conference (ICSEC),
2014 International. IEEE, 2014, pp. 163–167.
[24] B. Hongliang and L. Changping, “A hybrid license plate extraction
method based on edge statistics and morphology,” in Pattern Recogni-
tion, 2004. ICPR 2004. Proceedings of the 17th International Conference
on, vol. 2. IEEE, 2004, pp. 831–834.
[25] R. Wang, N. Sang, R. Wang, and L. Jiang, “Detection and tracking
strategy for license plate detection in video,” Optik-International Journal
for Light and Electron Optics, vol. 125, no. 10, pp. 2283–2288, 2014.
[26] H. Li and C. Shen, “Reading car license plates using deep convolutional
neural networks and LSTMS,” arXiv preprint arXiv:1601.05610, 2016.
[27] R. Laroca, E. Severo, L. A. Zanlorensi, L. S. Oliveira, G. R. Gonc alves,
W. R. Schwartz, and D. Menotti, “A robust real-time automatic li-
cense plate recognition based on the YOLO detector,” arXiv preprint
arXiv:1802.09567v2, 2018.
[28] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards realtime
object detection with region proposal networks,” in Adv. Neural Inf.
Process. Syst., 2015, p. 9199.
[29] M. H. Asghari and B. Jalali, “Edge detection in digital images using
dispersive phase stretch transform,” Journal of Biomedical Imaging, vol.
2015, p. 6, 2015.
[30] M. Sezgin and B. Sankur, “Survey over image thresholding techniques
and quantitative performance evaluation,” Journal of Electronic imaging,
vol. 13, no. 1, pp. 146–168, 2004.
[31] S. Zhu, S. Dianat, and L. K. Mestha, “End-to-end system of license plate
localization and recognition,” Journal of Electronic Imaging, vol. 24,
no. 2, pp. 023 020–023 020, 2015.
[32] P. Tarabek, “Fast license plate detection based on edge density and
integral edge image,” in Applied Machine Intelligence and Informatics
(SAMI), 2012 IEEE 10th International Symposium on. IEEE, 2012, pp.
37–40.
[33] “Medialab LPR database.” [Online]. Available:
http://www.medialab.ntua.gr/research/LPRdatabase.html