CLASSIFICATION AND COMPARISON OF LICENSE PLATES LOCALIZATION ALGORITHMS

Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.2, April 2021
DOI: 10.5121/sipij.2021.12201 1
CLASSIFICATION AND COMPARISON OF LICENSE
PLATES LOCALIZATION ALGORITHMS
Mustapha Saidallah, Fatimazahra Taki, Abdelbaki El Belrhiti El Alaoui and
Abdeslam El Fergougui
Computer Networks and Systems Laboratory, Faculty of Sciences, Moulay Ismail
University, Meknes, Morocco
ABSTRACT
The Intelligent Transportation Systems (ITS) are the subject of a world economic competition. They are the
application of new information and communication technologies in the transport sector, to make the
infrastructures more efficient, more reliable and more ecological. License Plates Recognition (LPR) is the
key module of these systems, in which the License Plate Localization (LPL) is the most important stage,
because it determines the speed and robustness of this module. Thus, during this step the algorithm must
process the image and overcome several constraints as climatic and lighting conditions, sensors and angles
variety, LPs’ no-standardization, and the real time processing. This paper presents a classification and
comparison of License Plates Localization (LPL) algorithms and describes the advantages, disadvantages
and improvements made by each of them.
KEYWORDS
License Plate Localization, LPL constraints, performance, real time, multi-localization, convolutional neural
network, deep learning.
1. INTRODUCTION
The mobility needs of people and products make the road traffic more and more cumbersome. ITS
are designed to circumvent the challenges engendered by this problem. LPR is the centerpiece of
this domain. It is based on three phases: the LPL, the characters segmentation of LP, and the Optical
Characters Recognition (OCR).
An efficient LPL algorithm should provide accurate results in real time. Therefore, it must be able
to manage the localization phase constraints: the multi-localization of all LP types, the noise due
to climatic conditions, dirty LP or complex acquisition scenes, the lighting conditions that cause
low contrast on the image or a multiple color appearance the non-standardization, the
sensor/vehicle distance and the capture angle that significantly influence on colors, sizes and shapes
of the LPs.
In our study, we classified the LPL algorithms into three families: the family of algorithms based
on mathematical processing is presented in section 2, the family of algorithms based on the plate
features is presented in section 3 and the family of algorithms based on deep learning in section 4.
Comparative mapping performance of the studied algorithms according to this phase constraints is
presented in section 5. In Section 6, a conclusion of our work will be provided.

2
2. FAMILY BASED ON MATHEMATICAL PROCESSING
The mathematical processing of images is a very active field, where the advanced theoretical are
concretized to manipulate the image content. It consists in transforming an image to another to
extract useful information for the application. It involves partial differential equations, geometric
tools, and optimization concepts. Then for the validation phase the LP properties are exploited.
This family can be divided into five algorithm classes based on: edges extraction, morphological
operations, hybrid methods, transforms and Sliding Concentric Windows (SCW).
2.1. Class of Algorithms based on Edge Extraction
These algorithms apply the convolution of high-pass filter and the original image; they highlight
the high frequencies and attenuate the rest. These are the techniques that transform the image into
a set of edges, not necessarily closed, forming the significant borders of image. These methods
consider the characters and the LP contours as a reference point for localization. They proceed
essentially in four steps:
• The Application of filter to extract the edge points.
• The binarization of the Contours image by choosing a threshold.
• The Analysis of result to extract the candidates.
• The validation by geometrical, textural or density properties.
The authors of [1] use the Sobel filter to extract edges from the color image, and the histogram to
select the binarization threshold. The logical AND of the binary and the HSV images, was exploited
to extract the yellow candidates, and the other types are extracted from the Integral Image. The
validation phase was based on the analysis of geometrical properties of Connected Component
(CC). The regions with the high vertical gradient density are the candidates on [2]. The validation
phase was based on aspect ratio, shape and tracking criterion.
In [3,4] the edge extraction is performed on the grayscale image to reduce the processing time and
the texture was the criterion used to exactly isolate the LP.
The edge extraction in [5,6] is performed by the Canny filter, it is a noise reduction by a Gaussian
filter before applying the Sobel filter. The contours density is the property that has confirmed the
LP region.
Simplicity and speed are the qualities of these methods. But these qualities are defects in case of
noisy images, complex acquisition scenes or if the LP borders do not present a great variation
compared to the rest of image. Moreover, these algorithms use two thresholds, the edge extraction
threshold which is difficult to be automatically selected under the various lighting conditions and
the region contour density threshold for the validation phase that makes these algorithms very
dependent to the sensor-vehicle distance.
2.2. Class of Algorithms based on Morphological Operations
This a set of nonlinear operators intended to probe the image with small geometries called
Structuring Elements (SE) to highlight certain shapes in the image.
• The choice of the adequate structuring element.
• The application of morphological operations based on Dilation and Erosion.

3
• A binary representation of the pixels (white present the objects and black present the
background).
• The CC Analysis based on the histogram of the results.
The authors of [7] describe a localization algorithm based on opening. It is a dilation of the erosed
grayscale image. The validation phase was based on the horizontal and vertical projection of the
result and the priori knowledge of the LP location. Therefore, the multilocalization is not allowed
by this algorithm.
In [8] a combination of hat transform and morphological gradient was used to extract the
candidates. The density, the location and the length/width ratio are used to confirm the LP region.
The strength of this class is its robustness to noise since it puts in advance the image shapes and
denoises it at a time. Nevertheless, its treatments take an enormous time what makes them
inadaptable to real-time systems. Moreover, the morphological operations don’t give good results
in case of poorly contrasted images where the sensitivity of these methods to the lighting
conditions. Finally, the choice of adequate SE is an important parameter influenced and limited by
a fixed sensor-vehicle distance.
2.3. Class of Hybrid Algorithms
These algorithms combine the edge detection and the morphological operations. They simplify the
image by keeping just edges. Then they apply morphological operations to denoise the picture and
complete the shapes, to define the CC. and based on geometrical and/or textural criteria the LPs
are validated.
In [9] the magnitude of the vertical gradients is used to detect candidate regions. The opening is
the morphological operation applied to reduce the noise and to remove the objects non-LP. Then
the length/width ratio, the size and the orientation of candidates are the geometrical criteria used to
validate the LP.
Treatment was reversed in [10] to solve the sensitivity of this class to lighting conditions. The
algorithm starts with a Top-hat and Bottom-hat transforms that allowed to achieve a high contrasted
grayscale image. Then the edges are extracted via the Sobel filter. The validation step consists in
the CC analysis according to geometric criteria.
The authors of [11,12] propose a candidate’s location by extracting the contours of the binary image
using the Canny filter. They added to the opening a dilation operation to connect close objects. The
region that checks the geometric criteria: rectangular shape and width/height ratio is a LP. Rotation
of the plate is carried out in [10] for correcting the deviation.
In [13] the extraction of the image contours is performed by the second derivative via the Laplacian
Of Gaussian (LOG) filter followed by a dilation to refine the contours and connect objects, and by
classifying candidates to "text" and "no text", the LP is exactly located.
These algorithms showed a great improvement of the localization rate. Thus, they lend well to the
no-standardization and the multi-localization constrain. However, the problems related to the edge
extraction (Identification of thresholds under different lighting conditions) and to the
morphological operations (Significant processing time) remain strongly posed.

4
2.4. Class of Algorithms based on Transforms
In this class we will study the sub-classes based on the Hough Transform (HT), the Wavelet
Transform (WT) and Generalized Symmetry Transform (GST).
The Hough Transform is a technique of simple pattern recognition in an image by binding the
pixels between them; its major disadvantage is the great processing time. To solve this problem,
the authors of [14] proposed a pretreatment step which consists in binarizing the contour map of
the original image, and on the result they applied the Hough Transform to extract the parallelogram
objects. The validation phase consists in assessing the length/width ratio of candidates, then
calculating the number of objects cut by two horizontal lines drawn at 1/3 and 2/3 of the candidate’s
height.
The HT does not give a good result when the image is noisy. The authors of [15] proposed to
denoise the image via a Median filter to solve this problem. Then they applied a second filter to
isolate the area where the plate is probably located to reduce the processing time.
The results of this subclass are accurate, and it can locate any type of plates at once, because it
focuses on the shape of the plate and not on the type. However, these algorithms cannot be
implemented in real-time systems or on images where the rectangular shape of the LP is not clear
(Because of an obscuring object, continuity of intensity, or physical damage). Further, this type of
algorithms cannot locate too deviated plates.
The Wavelet Transform decomposes the image to a set of cards, each presenting a characteristic.
The authors of [16,17] started from the idea that the great difference in intensity between the LP
and its characters is a key element to locate it. Therefore, they decompose the grayscale image into
four frames via the WT, of which they used two. On the first, containing the horizontal details, they
search the area where there is a maximum value of variation to locate a reference line. By analyzing
the horizontal projection curve of the second image, containing the vertical details, they arrived to
determine the sizes of the existing LPs on the image. Then by scanning the image they determined
candidates. The validation phase was based on the geometric properties analysis.
These algorithms are independent of the sensor-vehicle distance and they are capable to detect
several LP at the same time. However, they give false results when the contrast between the
characters and the background is not quite clear. Thus, their sensitivity to the noises is high.
The GST in [18] evaluates evidence of local symmetry; it extracts a map of symmetrical contours
of the image in all directions. The authors used in [19] the symmetry of the LP like an element-key
to locate it. And to reduce the processing time the authors focused on the extraction of the four
corners of the plate; consequently, they scanned the image only in the two diagonal directions (45
and -45). The multiplication of the two resultants and the connection of the extracted corners, give
the LP candidates. Finally, a neural network system is used to validate the LP.
To adapt this subclass to real-time systems it is necessary to fix a limited number of scans. Thus,
to achieve accurate results it is necessary to consider the lighting conditions and noise constraints.
Moreover, the algorithm becomes weakening in the case of damaged LPs or if one of its corners is
hidden.
2.5. Class of Algorithms based on Sliding Concentric Windows
This method is based on pixel intensities to extract the regions of interest. The LP localization in
[20] passes by two stages: The first consists in scanning the image by two SCW using the standard

5
deviation; if those values exceeds the predefined thresholds, the pixel belongs to a region of interest
and it is posted at 1, if not the pixel is worth 0. The result is a binary image of candidates. The
second phase uses geometrical criteria to validate the LP. The method is adapted to the multi-
localization applications. However, since it is based on the pixel’s intensities, a low-resolution or
a poor lighting conditions are its blocking points. Thus, its simplicity made it unable to process
noisy images or complex scenes.
3. FAMILY BASED ON LP FEATURES
These methods assume that a good prior knowledge of the LP properties, help to implement a
robust model of its localization. They consider that the LP is an object that has defining properties
in the image: specific texture, unique color composition, or both combinations. Identifying the LP
properties is the first step of these algorithms. Then, each pixel is represented by a vector of these
properties. Therefore, the classification methods classify and label pixels to indicate their affiliation
or not to a region. Finally, the validation phase is based on geometrical criteria or histogram
analysis. The localization rate of these methods is generally high, but they are specific to the treated
LPs types.
3.1. Class of Algorithms based on LP Texture
The algorithms of this class are focused on the analysis of LP texture. They assume that the LP and
its characters have a specific texture that distinguishes them from other image objects.
Based on the four following characteristics:
• In a small region contains more than one character,
• Characters and background are in sharp contrast,
• The size of the plate region is relatively fixed with a fixed length and width ratio,
• The tilt angle of license plate is in a certain range.
The algorithm in [21] constructs the attribute vectors to classify the pixels. The validation phase
was based on the analysis of the result histogram.
The idea of [22] is that the LP is in a textural region of the image. Therefore, they decompose the
image into blocks. The density of each block is calculated to classify it in "textural" or "not
textural". The validation phase was based on geometrical criteria.
The authors of [23] implement four LPL algorithms based on the following LP features:
• The length/width ratio of the connected region,
• The background area,
• The character/plate area density,
• The number of edges in the plate center.
These features are exploited by deterministic, probabilistic and both combination methods, taking
several thresholds for each one to validate the results.
This class uses the values showing the properties degree in a region. These values are directly
affected by the noise, the sensor-vehicle distance, and the image resolution.

6
3.2. Class of Algorithms based on LP Colors
A unique colors combination which occurs almost exclusively in the LP area is the basic idea of
these algorithms. In [24] the pixels classification is performed in the HSV (Hue, Saturation, Value)
space to solve the problem of the wide RGB (Red, Green, Blue) value under different illumination.
The validation phase was based on the evaluation of the length/width ratio and the projection curve
analysis.
The authors of [25] have based their method on Fuzzy Logic. They decompose the color image
into 4 cards: E denote the edge map computed using the color edge detector implemented by
authors, H, S and V are the maps preserving the hue, saturation, and intensity components of the
transformed image. E, H, S and V are the entries of the fuzzy system that classified the pixels. The
result represents the degrees of the pixels belonging to a LP. The validation phase was delegated
to the OCR. The richness of the test base confirmed the algorithm localization performance.
However, the processing time made it inapplicable in real time systems.
The algorithms of this class reach high localization rates, but in an important processing time.
Moreover, the sensitivity to the lighting conditions is their major disadvantage. Furthermore, the
LPs have different colors and shapes because of the non-standardization which make these
algorithms specifics to treated LP types.
3.3. Class of Algorithms based on Vector Quantization (VQ)
This concept is described in [26]. It is to define a set of vectors forming a dictionary. They include
information on the image blocks. These blocks are obtained by selecting the high contrasted strips
in a quad tree decomposition phase. These strips are validated by sorting them according to the
expected LP size and the code words that describe the blocks content.
The authors of [26] assume that the LP corresponds generally to the first two bands of the sorted
list. The test of this method in a real system has shown its high dependence to the sensor-vehicle
distance (A considerable change in this factor requires the update of the entire system). Moreover,
it can’t process too noisy images or localize several plates at a time.
4. FAMILY BASED ON DEEP LEARNING
Deep learning is a new technique in the area of machine learning, which attempts to model high-
level abstractions in data. There are various deep learning architectures such as deep convolutional
neural network, deep belief network, and so on. Due to the deep network structures, they have been
widely used in many applications with great success. Compared with other architectures, a deep
convolutional neural network has fewer weights and less complexity. In addition, it allows one to
directly use original images as inputs which enable a hierarchical learning of features.
The authors of [27] proposed a work-flow and methods for segmentation and annotation free
Automated LPR (ALPR) with improved plate localization and automatic failure identification.
Their proposed workflow first localizes the license plate region in the captured image using a two-
stage approach where a set of candidate regions are first extracted using a weak SNoW classifier
and then scrutinized by a strong convolutional neural network (CNN) classifier in the second stage.
Images that fail a primary confidence test for plate localization are further classified to identify
reasons for failure, such as license plate not present, LP too bright, LP too dark or no vehicle found.
In the localized plate region, they perform segmentation and OCR jointly by using a probabilistic
inference method based on hidden Markov models where the most likely code sequence is

7
determined by applying the Viterbi algorithm. In order to reduce manual annotation required for
training classifiers for OCR, they propose to use either artificially generated synthetic license plate
images or character samples acquired by trained ALPR systems already operating in other sites.
The performance gap due to differences between training and target domain distributions is
minimized using an unsupervised domain adaptation. They evaluated the performance of their
proposed methods on license plate images captured in several US jurisdictions under realistic
conditions.
To avoid large accumulated error in typical three-step LPR methods and improve recognition
performance in challenging complex environment, an end-to-end recognition method based on
deep convolution neural network named LPR-Net is proposed in [28]. The input of LPR-Net is a
gray or color image with any size and the output is the license plate number of input image. If no
license plate in input images, LPR-net will output “no license”. The proposed LPR-Net is a hybrid
deep architecture that consists of a basic net, a multi-scale net, a regression net, and a classification
net. Firstly, LPR-Net resizes the input image to 500 × 500 pixels. Then the resized image is input
to basic net to get basic features. After that, the basic features are input to multiscale net to obtain
multi-scale features. Finally, the multi-scale features are input to regression net and classification
net to locate plate and identify characters.
The authors of [29] developed a deep learning application to detect and recognize Korean cars’
license plates from images. It is an advanced application that targets to provide deep learning
solution that can be applied in many areas including Intelligent Transportation System, Internet of
Things and Smart City. They developed their method that is a combination of scene text recognition
technique with Geometrical Image Transformation (GIT) to recognize number plates for combined
neural networks and achieving 99.8% and 95.7% of detection and recognition accuracy
respectively.
The authors of [30] proposed a novel machine learning approach to recognizing the license plate
number. They used one of the most successful deep learning methods, convolutional neural
network for extracting the visual features automatically. Suitable localization and segmentation
techniques are employed before CNN model to enhance the accuracy of the proposed model. In
addition to this, the Deep License Plate Number Recognition (D-PNR) model also takes care of
proper identification from images those are hazy and is not suitable-inclined or noisy images. The
efficiency of license plate recognition system improved using deep Learning framework to keep
the trade-off between accuracy and time complexity. So, an efficient and real-time-based D-PNR
model proposed for the license plate recrimination. It is able to give end results with 98.02%
accuracy after setting suitable learning parameters.
The authors of [31] presented an embedded platform-based Italian license plate detection and
recognition system using deep neural classifiers. In their work, trained parameters of a highly
precise ALPR system are imported and used to replicate the same neural classifiers on an NVidia
Shield K1 tablet. A CUDA-based framework is used to realize these neural networks. The flow of
the trained architecture is simplified to perform the license plate recognition in realtime. Results
show that the tasks of plate and character detection and localization can be performed in real-time
on a mobile platform by simplifying the flow of the trained architecture. However, the accuracy of
the simplified architecture would be decreased accordingly.
The authors of [32] used deep learning method for real-time traffic images. They first extract
license plate candidates from each frame using edge information and geometrical properties,
ensuring high recall using one class Support Vector Machines (SVM). The verified candidates are
used for Number Plate (NP) recognition purpose along with a spatial transformer network (STN)

8
for character recognition. Their system is evaluated on several traffic images with vehicles having
different license plate formats in terms of tilt, distances, colors, illumination, character size,
thickness etc. Also, the background was very challenging. Results demonstrate robustness to such
variations and impressive performance in both the localization and recognition.
5. COMPARISON AND SYNTHESIS
The exactitude of a LPR system depends on the localization phase, i.e., to be efficient, its
localization rate should be around 100%. The experimental tests of localization algorithms are not
done on unified databases. Indeed, a rich database that allows to test the performance of the
algorithms must contain many several plate images, taken in different lighting conditions, from
various distances, angles, and acquisition scenes. These images must present all LP types. In
addition, a localization algorithm should not simply isolate the LP, but it must also isolate it in real
time.
To qualify the algorithms, we distinguish four levels of the “performance” parameter. This later
binds the localization rate and the database richness as the Table 1 shows.
Indeed, if the database contains a significant number of images and various constraints related to
this phase, its richness is very high and if the localization rate is around 100%, then the algorithm
will be qualified “very powerful: AAA”.
Table 1. Values of the performance parameter.
Database richness Localization rate Performance
Very high Very high AAA
Very high Medium AA
Medium High AA
Medium Medium A
Low High A
High Low A
Low Medium B
Medium Low B
Low Low B
In addition to the performance of the LPL algorithms, Table 2 shows their sensitivity to the
localization phase constraints. This sensitivity is deduced from the analysis of algorithms and
which is not necessarily present in the database. For example, if the database does not contain
images of vehicles with constraints:
• LP type: the database contains only the car LP of a single line,
• No-standardization: the database contains only the plates respecting a given standard.
And if the analysis shows that the algorithm can support only one LP type of different standards,
the answer in Table 2 will be "No" for the LP type and "Yes" for the Nostandardization.
From Table 2, we deduce on the one hand, the absence of an ideal LPL algorithm: very powerful
and which solve the various constraints of this phase. Moreover, the most efficient algorithms are

9
deep learning algorithms because they solve most constraints. On the other hand, the least
overcome constraint by the different algorithms is sensitivity to lighting conditions. Thus, this
constraint is circumvented just by nine algorithms; three of them are based on the LP features and
four on deep learning.
Table 2. Performance of LPL algorithms and their sensitivity to localization constraints.
6. CONCLUSIONS
We have presented the functioning principle of different LP Localization methods by analyzing the
strengths and weaknesses of each of them. Our study has demonstrated the lack of an ideal

10
localization algorithm and confirmed that a meaningful comparison of these methods performance
requires their evaluation on the same test database, containing enough scenes with the various
constraints of this phase. Finally, we fill a table showing main properties of the considered
algorithms which could help a researcher to select the most appropriate one for his specific task
and conditions.
REFERENCES
[1] L. Luo, H. Sun, W. Zhou, and L. Luo, “An Efficient Method of License Plate Location,” in 2009 First
International Conference on Information Science and Engineering, Dec. 2009, pp. 770–773, doi:
10.1109/ICISE.2009.250.
[2] N. Thome, A. Vacavant, L. Robinault, and S. Miguet, “A cognitive and video-based approach for
multinational License Plate Recognition,” Machine Vision and Applications, vol. 22, no. 2, pp. 389–
407, Mar. 2011, doi: 10.1007/s00138-010-0246-3.
[3] K. Parasuraman and P. V. Kumar, “An efficient method for indian vehicle license plate extraction and
character segmentation,” in IEEE International conference on computational intelligence and
computing research, 2010, vol. 18, pp. 1–4.
[4] P. Tarabek, “A real-time license plate localization method based on vertical edge analysis,” in 2012
Federated Conference on Computer Science and Information Systems (FedCSIS), Sep. 2012, pp. 149–
154.
[5] S. Draghici, “A neural network based artificial vision system for licence plate recognition,”
International Journal of Neural Systems, vol. 8, no. 01, pp. 113–126, 1997.
[6] K. P. Thooyamani, V. Khanaa, and R. Udayakumar, “Application of pattern recognition for farsi license
plate recognition,” Middle-East Journal of Scientific Research, vol. 18, no. 12, pp. 1768– 1774, 2013.
[7] C. Wu, L. C. On, C. H. Weng, T. S. Kuan, and K. Ng, “A Macao license plate recognition system,” in
2005 International Conference on Machine Learning and Cybernetics, 2005, vol. 7, pp. 4506– 4510.
[8] H. K. Sulehria, Y. Zhang, and D. Irfan, “Mathematical morphology methodology for extraction of
vehicle number plates,” International journal of computers, vol. 1, no. 3, pp. 69–73, 2007.
[9] Shen-Zheng Wang and Hsi-Jian Lee, “Detection and recognition of license plate characters with
different appearances,” in Proceedings of the 2003 IEEE International Conference on Intelligent
Transportation Systems, Oct. 2003, vol. 2, pp. 979–984 vol.2, doi: 10.1109/ITSC.2003.1252632.
[10] A. Rabee and I. Barhumi, “License plate detection and recognition in complex scenes using
mathematical morphology and support vector machines,” in IWSSIP 2014 Proceedings, May 2014, pp.
59–62.
[11] V. D. Mai, D. Miao, R. Wang, and H. Zhang, “An Improved Method for Vietnam License Plate
Location, Segmentation and Recognition,” in 2011 International Conference on Computational and
Information Sciences, Oct. 2011, pp. 212–215, doi: 10.1109/ICCIS.2011.79.
[12] Bai Hongliang and Liu Changping, “A hybrid license plate extraction method based on edge statistics
and morphology,” in Proceedings of the 17th International Conference on Pattern Recognition, 2004.
ICPR 2004., Aug. 2004, vol. 2, pp. 831-834 Vol.2, doi: 10.1109/ICPR.2004.1334387.
[13] R. Radha and C. P. Sumathi, “A novel approach to extract text from license plate of vehicles,” Signal
& Image Processing, vol. 3, no. 4, p. 181, 2012.
[14] T. D. Duan, T. H. Du, T. V. Phuoc, and N. V. Hoang, “Building an automatic vehicle license plate
recognition system,” in Proc. Int. Conf. Comput. Sci. RIVF, 2005, vol. 1, pp. 59–63.
[15] P. Prabhakar, P. Anupama, and S. R. Resmi, “Automatic vehicle number plate detection and
recognition,” in 2014 International Conference on Control, Instrumentation, Communication and
Computational Technologies (ICCICCT), Jul. 2014, pp. 185–190, doi:
10.1109/ICCICCT.2014.6992954.
[16] Ching-Tang Hsieh, Yu-Shan Juan, and Kuo-Ming Hung, “Multiple license plate detection for complex
background,” in 19th International Conference on Advanced Information Networking and Applications
(AINA’05) Volume 1 (AINA papers), Mar. 2005, vol. 2, pp. 389–392 vol.2, doi:
10.1109/AINA.2005.257.
[17] A. Boudjella, B. B. Samir, H. B. Daud, and R. Syahira, “License plate recognition part II: Wavelet
transform and Euclidean distance method,” in 2012 4th International Conference on Intelligent and
Advanced Systems (ICIAS2012), Jun. 2012, vol. 2, pp. 695–700, doi: 10.1109/ICIAS.2012.6306103.

11
[18] D. Reisfeld, H. Wolfson, and Y. Yeshurun, “Context-free attentional operators: The generalized
symmetry transform,” International Journal of Computer Vision, vol. 14, no. 2, pp. 119–130, Mar.
1995, doi: 10.1007/BF01418978.
[19] Dong-Su Kim and Sung-Il Chien, “Automatic car license plate extraction using modified generalized
symmetry transform and image warping,” in ISIE 2001. 2001 IEEE International Symposium on
Industrial Electronics Proceedings (Cat. No.01TH8570), Jun. 2001, vol. 3, pp. 2022–2027 vol.3, doi:
10.1109/ISIE.2001.932025.
[20] C. Anagnostopoulos, T. Alexandropoulos, S. Boutas, V. Loumos, and E. Kayafas, “A templateguided
approach to vehicle surveillance and access control,” in IEEE Conference on Advanced Video and
Signal Based Surveillance, 2005., Sep. 2005, pp. 534–539, doi: 10.1109/AVSS.2005.1577325.
[21] F. Jun and D. Shuguang, “A vehicle license plate location and correction method based the
characteristics of license plate,” in Proceedings of the 10th World Congress on Intelligent Control and
Automation, Jul. 2012, pp. 42–46, doi: 10.1109/WCICA.2012.6357836.
[22] H. Anoual, S. EL FKIHI, A. JILBAB, and D. ABOUTAJDINE, “A NOVEL TEXTURE-BASED
ALGORITHM FOR LOCALIZING VEHICLE LICENSE PLATES.,” Journal of Theoretical &
Applied Information Technology, vol. 45, no. 1, 2012.
[23] R. Al-Hmouz and S. Challa, “License plate localization based on a probabilistic model,” Machine
Vision and Applications, vol. 21, no. 3, pp. 319–330, Apr. 2010, doi: 10.1007/s00138-008-0164-9.
[24] X. Shi, W. Zhao, and Y. Shen, “Automatic License Plate Recognition System Based on Color Image
Processing,” in Computational Science and Its Applications – ICCSA 2005, Berlin, Heidelberg, 2005,
pp. 1159–1168, doi: 10.1007/11424925_121.
[25] Shyang-Lih Chang, Li-Shien Chen, Yun-Chung Chung, and Sei-Wan Chen, “Automatic license plate
recognition,” IEEE Transactions on Intelligent Transportation Systems, vol. 5, no. 1, pp. 42–53, Mar.
2004, doi: 10.1109/TITS.2004.825086.
[26] 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, Feb. 2000, doi: 10.1109/41.824138.
[27] O. Bulan, V. Kozitsky, P. Ramesh, and M. Shreve, “Segmentation- and Annotation-Free License Plate
Recognition With Deep Localization and Failure Identification,” IEEE Transactions on Intelligent
Transportation Systems, vol. 18, no. 9, pp. 2351–2363, Sep. 2017, doi: 10.1109/TITS.2016.2639020.
[28] D. Wang, Y. Tian, W. Geng, L. Zhao, and C. Gong, “LPR-Net: Recognizing Chinese license plate in
complex environments,” Pattern Recognition Letters, vol. 130, pp. 148–156, Feb. 2020, doi:
10.1016/j.patrec.2018.09.026.
[29] S. Usmankhujaev, S. Lee, and J. Kwon, “Korean License Plate Recognition System Using Combined
Neural Networks,” in Distributed Computing and Artificial Intelligence, 16th International Conference,
Cham, 2020, pp. 10–17, doi: 10.1007/978-3-030-23887-2_2.
[30] K. Kumar, S. Sinha, and P. Manupriya, “D-PNR: Deep License Plate Number Recognition,” in
Proceedings of 2nd International Conference on Computer Vision & Image Processing, Singapore,
2018, pp. 37–46, doi: 10.1007/978-981-10-7898-9_4.
[31] S. T. H. Rizvi, D. Patti, T. Björklund, G. Cabodi, and G. Francini, “Deep Classifiers-Based License
Plate Detection, Localization and Recognition on GPU-Powered Mobile Platform,” Future Internet,
vol. 9, no. 4, Art. no. 4, Dec. 2017, doi: 10.3390/fi9040066.
[32] S. Sharma, P. Kumar, and K. Kumar, “A-PNR: Automatic Plate Number Recognition,” in Proceedings
of the 7th International Conference on Computer and Communication Technology, New York, NY,
USA, Nov. 2017, pp. 106–110, doi: 10.1145/3154979.3154999.

CLASSIFICATION AND COMPARISON OF LICENSE PLATES LOCALIZATION ALGORITHMS

Recommended

More Related Content

What's hot (20)

Similar to CLASSIFICATION AND COMPARISON OF LICENSE PLATES LOCALIZATION ALGORITHMS (20)

Recently uploaded (20)

CLASSIFICATION AND COMPARISON OF LICENSE PLATES LOCALIZATION ALGORITHMS