Journal of Engineering and Technology for Industrial Applications

ITEGAM-JETIA
Manaus, v.10 n.48, p. 129-134. July/August., 2024.
DOI: https://doi.org/10.5935/jetia.v10i48.955
ISSN ONLINE: 2447-0228

RESEARCH ARTICLE OPEN ACCESS

AUTOMATIC LICENSE PLATE RECOGNITION SYSTEM: A SYSTEMATIC

SURVEY
Vishakha H. Jagtap1, Rohit V. Dhotre2, Utkarsh R. Khandare3, Harshada N. Khuspe4, Rohini B.
Kokare5
1,2,3,4 Student, Department of Artificial Intelligence and Data Science, VPKBIET, Baramati, 413133, India
5Assistant Professor, Department of Artificial Intelligence and Data Science, VPKBIET, Baramati, 413133, India
1http://orcid.org/0009-0008-0590-2499 , 2http://orcid.org/0009-0001-0584-2245 , 3http://orcid.org/0009-0006-2494-071X ,
4http://orcid.org/0009-0000-6527-692X , 5 http://orcid.org/0009-0002-1633-1317

Email: [email protected], [email protected], [email protected], [email protected],

[email protected]

ARTICLE INFO ABSTRACT

Article History The capacity to naturally distinguish and extract License plate data from pictures or video
Received: November 16 th , 2023 streams has gathered noteworthy consideration in later a long time, owing to its potential to
Revised: July 08 th, 2024 upgrade security, streamline activity operations, and encourage effective information
Accepted: July 18th , 2024 collection. The Vehicle Number Plate Recognition (VNPR) system has a broader variety of
Published: August 30th, 2024 applications. A sophisticated License Plate Recognition (LPR) system can be smoothly
incorporated into existing processes including law enforcement, monitoring, and toll station
Keywords: services. Existing approaches for License Plate Recognition are limited to datasets like
Automatic license plate CCPD, AOLP, etc., and operation specific, so many of them require a constrained
recognition (ALPR), environment to meet the needs of the intended application. Even if there are many Vehicle
Authentication, Number Plate Recognition systems available today, the task is still difficult because of
Character recognition, several aspects such as the fast-moving vehicles, inconsistent vehicle number plates,
Character segmentation. contrast problems, language of the vehicle number, processing and memory limitations,
camera mount position, motion-blur, reflections, tolerance to distortion, and varying lighting
conditions. The methodologies and procedures employed for ALPR in Deep Learning,
Computer Vision, and Machine Learning domains in contemporary literature are
investigated in this study. This paper gives a comparative study of the techniques and
algorithms used for various tasks included in Vehicle Number Plate Recognition (VNPR)
systems such as License Plate Detection, License Plate Recognition, Character
Segmentation, etc. We outline a critical and constructive analysis of relevant studies in the
ALPR, and it will also give directions for future research, and optimization of the current
approaches.

Copyright ©2024 by authors and Galileo Institute of Technology and Education of the Amazon (ITEGAM). This work is licensed
under the Creative Commons Attribution International License (CC BY 4.0).

I. INTRODUCTION for an automatic license plate recognition system. The techniques

are as follows: [1] Many techniques are used for license plate
I.1 OVERVIEW recognition, but mostly CNN technology is the most accurate
Currently, the automatic number plate recognition system is technique for license plate recognition of a vehicle using a camera
very useful in many fields to identify cars or vehicles, which can placed on the road or public places. In [2] paper the technique used
be classified by the number plate with any background color, which to detect license plate number is faster RCNN. This technique is
helps to identify the type of vehicle. It helps a lot. to many people more accurate than CNN. Using this faster RCNN model, we can
who work in RTO or identify the vehicle and which car is illegal detect any tile with any background color or number of font styles.
or legal This system has been used in many countries to control For [3] Next technique is the OpenCV technique. This technique is
traffic laws and to control urban traffic. Many techniques are used based on Computer Vision. The next techniques are Vsnet [4],

Journal homepage: www.itegam-jetia.org

 One, Two and Three, ITEGAM-JETIA, Manaus, v.10 n.48, p. 129-134, July/August., 2024.

Support Vector Machine, RCNN and Cascade Color Space II. AUTOMATED LICENSE PLATE RECOGNITION
Transformation of Pixel Features, Cascaded Contrast-Color Haar- APPROACHES
like Features, Cascaded Convolution Network. The Multi-Level
Extended Local Binary Patterns and Extreme Learning Machine II.1 MULTI-STAGE LICENSE PLATE RECOGNITION
are also used to detect the number plate from vehicles. SYSTEMS
License plate recognition approaches can be broadly
I.2 SURVEY MOTIVATION classified as multi-stage and single stage approaches. In multistage
Although automatic number plate recognition systems are approaches mainly three classes are there. These include License
intended for outdoor usage, they struggle to find and identify plate detection or extraction [1], License Plate segmentation
license plates in constantly changing weather and environmental (extracting separate characters), and character recognition.
circumstances. The application of most current systems is In the first stage, License plate is detected by using an
constrained by elements like shifting lighting conditions, snow or attention mechanism. Region proposal network (RPN) generates
fog, day and night, camera shaking, rotations, and occlusions. the rectangular object proposals for the subsequent processing that
Automatic Number Plate Recognition systems, which are sensitive helps to detect the number plate efficiently [1],[2],[5]. The
to changes in light and typically work in daylight, must deal with cascaded CC-Haar-like detector, the cascaded CST-pixel detector,
cars traveling at varied speeds in the real world. Many methods, and the cascaded ConvNet detector make up a hybrid cascade for
which primarily function in daylight, are sensitive to variations in the detection of license plates with various resolutions [3]. A
light. Production Systems for automatic number plate recognition vertexNET architecture, composed of two cascaded CNNs, is used
must also satisfy non-functional criteria including acquisition and for license plate recognition [6]. The license plate detection stage
operation costs, physical specifications, power needs, and is classified into two stages namely feature extraction stage and
connection restrictions. ELM classification stage [7].
In the second stage, some common techniques for the
I.3 ARTICLE STRUCTURE segmentation and individual character extraction are used such as
vertical or horizontal projection [1]. There are some algorithms in
A comprehensive overview of license plate recognition the ALPR system that don't make use of segmentation techniques.
systems is provided in Section 1. The two basic strategies are So, this is not a mandatory stage in the ALPR system.
covered in Section 2. Multi-stage and single-stage ALPR for In the last stage, the character recognition algorithms are
license plate recognition. Principal components of a multi-stage used such as OCR or neural networks for character recognition.
license plate identification, such as license plate detection and Every stage in the multi-stage approach is equally important for the
license plate recognition are discussed in Sections 3 and 4. These overall performance of the ALPR system.
sections each provide the relevant advantages, difficulties,
restrictions, and suggested solutions. In Section 5 different datasets
used for the study are discussed with their respective parameters,
advantages, and disadvantages. In Section 6, we quickly go over
several evaluation parameters for the ALPR system. Section 7
addresses the challenges that need to be addressed for optimal
performance of the ALPR system. Finally, Section 8 concludes the
study.

Figure 2: Stages in multi-stage ALPR system.

Source: Authors, (2024).

II.2 SINGLE-STAGE LICENSE PLATE RECOGNITION

SYSTEMS
In the single stage approach, all the processes that are
separated in the multi-stage approach are addressed at a single
stage. Most of the advances are made in multi-stage approaches but
some do use single stage approaches. These all use a single deep
neural network that has been trained to detect, locate, and identify
the license plate from beginning to end in a single forward pass.
Such an attempt was made in the paper by Chen et al. The popular
VGG-16 [4] was utilized as the basis for the backbone network he
employed, which kept the convolutional layers from conv1_1 to
conv5_3. For a fair comparison, the final two fully connected
layers (fc6, fc7) are turned into convolutional layers and additional
layers from conv6_2 to conv9_2 are also included for semantically
stronger feature extraction. According to [4] proposed a 30
convolutional layers architecture structured into 9 modules, where
Figure 1: Structure of Paper. all of them have linear residual connections except for the first and
Source: Authors, (2024). the last one. Using interleaved separable convolutions and max
pooling, the entry flow raises the feature channel from 3 to 256

Page 130
 One, Two and Three, ITEGAM-JETIA, Manaus, v.10 n.48, p. 129-134, July/August., 2024.

while down sampling the spatial dimension from 160 48 to 40 6. III.2 YOLO-v2
They use repeated ResSeparableConv blocks in the middle flow,
keeping the spatial size and channel number constant, to extract The deep learning model YOLOv2 (You Only Look Once
deep features that contain higher level representations. We extract version 2) is frequently used for tasks involving the identification
a middle-level feature map M of size 406512 as the attention of license plates. YOLOv2, which has been trained on datasets with
network's context from the exit flow, along with a final feature annotated license plate areas, is excellent at real-time object
vector F of 512 dimensions. In the sequence decoder, LSTMs with identification. When used on license plates, it successfully
two layers and 512 hidden states each are used [8]. A 2-layer recognizes and localizes them inside picture or video frames. It is
attention mechanism is used which reduces the need for a useful tool for many applications since the model forecasts
segmenting each character in the license plate separately. A similar bounding boxes that include the precise placements of the license
kind of single stage approach is used in which use of inception v3 plates. A YOLOv2 [8] detector is utilized to obtain the bounding
with three layers of CNN and six layers of SSD 300 are used [9]. boxes of license plates.

III.3 VERTEXNET
VertexNet [6] is a good performing one-stage detector with
a limited input size, a narrow channel of high-level layers, and
vertex estimation. Three components, the head, fusion, and
backbone networks, make up the proposed VertexNet [6].
VertexNet is built using small-resolution input, even with character
information lost, to achieve fast inference speed and reduce
memory use [6].

III.4 DEPTH WISE SEPARABLE CONVOLUTIONAL

BLOCK (DSCB)
A Depth Wise Separable Convolution block (DSCB) is a
frequent building element in convolutional neural networks
(CNNs), particularly in topologies built for efficient and
lightweight model architectures such as MobileNet and Xception.
Depth wise convolution and pointwise convolution are the two
primary parts of the DSCB. [5] The Depth Wise Separable
Convolution block (DSCB) is used to reduce the number of
parameters to a minimum, allowing the model to be deployed on
mobile devices.

Figure 3: End-to-end CNN architecture for LP detection. III.5 EDGE DETECTION

Source: [8].
In computer vision and image processing, edge detection is
III. LICENSE PLATE DETECTION a key idea. It describes the method of locating the areas of a picture
where substantial changes in intensity or color take place. Usually,
A crucial part of contemporary transportation and security these transitions depict the edges of features or objects in the
systems is license plate detection. It involves locating and picture. Most of the research has relied on edge-based techniques
identifying license plates on automobiles for a variety of purposes. for license plate identification since every license plate is
Detection techniques have evolved from conventional computer rectangular and has a defined aspect ratio. A segmentation
vision methods to more recent advances in deep learning. Finding technique built on edge detection was utilized in [10]. To find the
the location of the vehicle license plate from the ingested vehicle license plate, it primarily employs the horizontal projection
picture and accurately segmenting the license plate from the area approach and the vertical projection method. The license plates
for character segmentation constitute the primary tasks of the approximate contour is found using Sobel edge detection [10].
license plate location. As a result, one of the crucial elements that
affects how well the system works is the choice of the license plate IV. LICENSE PLATE DETECTION
area. Mentioned below are license plate detection approaches.
License plate recognition in ALPR systems can be broadly
III.1 HYBRID CASCADE classified into three classes, pre-processing, character
segmentation, and character recognition.
A specific design is utilized to increase the precision and
effectiveness of license plate detecting systems called a hybrid IV.1 PRE-PROCESSING
cascade structure. To accomplish robust and accurate license plate
localisation, it integrates components of both a cascade classifier The main purpose of pre-processing is to enhance the image
and deep learning methods. The detection of license plates with quality to correctly enable the character recognition and henceforth
various resolutions is done using a hybrid cascade [3]. The enhance the recognition process. Pre-processing helps to neglect
cascaded CST-pixel detector, cascaded ConvNet detector, and the noise, prevent the image blur, etc. The effectiveness of
cascaded CC-Haar-like detector make up this hybrid cascade recognition will be impacted by the color information in color
structure's three components [3]. images. The license plate information is retained when the color

Page 131
 One, Two and Three, ITEGAM-JETIA, Manaus, v.10 n.48, p. 129-134, July/August., 2024.

image is converted to a grayscale image, and processing speed is has a total of 1376 photographs taken from the Hong Kong-Zhuhai-
also considerably increased [11]. Macao Bridge's entrance control system. The toll gate is where the
The License Plate is resampled and rectified to a higher test photographs are taken. There are 280 LPs photos altogether
resolution (64 * 256) according to the vertices. This resampling with a resolution of 1024 × 800 in the testing subset. Vehicles from
step normalizes the location of characters in the LP. Another Hong Kong, Mainland China, and Macao with a maximum of three
process performed is the rectification, which is achieved using license plates are included in the dataset.
perspective transformation. This generates the bird’s eye view of
the detected license plate [6]. Multi-level pre-processing V.2 ALOP DATASET
approaches are used to pass the detected license plate through
multiple stages of pre-processing. A Gaussian filter and the The 2049 license plate images in the ALOP [1],[6]
CLAHE method are used for multi-level pre-processing in [7],[9]. collection are broken up into three categories: road patrol (RP),
In this paper, G (σ) is the Gaussian filter with the standard deviation traffic law enforcement (LE), and access control (AC). The photos
σ = 0.25. CLAHE (σ) is defined as a contrast-limited adaptive of the RP subgroup are taken at different distances and
histogram equalization method with a standard deviation σ = perspectives. The AOLP dataset is split into two subsets: training
0.01[7]. A new image is produced at each stage of multi-level pre- pictures are utilized for the remaining photos, and files with
processing which eventually expands the training image size. filenames that begin with "1" are used as testing images (111
Based on the results of the LP detection, the License Plate images images). The ratio between the groups for training and testing is
are cropped depending on the required ROI. These cropped images around 4.5:1. LPs and characters are made and annotated using the
can be tilted vertically and horizontally. In [5] the focus is on the AOLP dataset's ground truth data.
vertical tilt only. The image is binarized after correction of
horizontal tilt. Selecting the specific pixels, starting from the first V.3 PKU DATASET
pixel value to judge the tilt level [8].
The PKU collection contains 3977 photos with Mainland
IV.2 CHARACTER SEGMENTATION China LPs. Because it solely provides the ground-truth file of LPs,
this dataset is used to assess the efficacy of LP detection. Five
The Mask branch is used for instantly segmenting the LP groups (G1–G5) make up the PKU dataset [1], with G1 being the
characters provided input of Region Proposals which is output from simplest and G5 being the most challenging. The [1] utilizes the
the Faster R-CNN [1]. The recognition of the rivet position and other 4 groups as the testing datasets and 810 photos from G1 as
white dots on the license plate to perform character segmentation the training dataset. PKU Data [6] is an LP detection dataset in
[10]. which the characters in 2253 images are labelled. Three subsets are
used to choose those images: G1 (daytime under normal
IV.3 CHARACTER RECOGNITION conditions), G2 (daytime with sun glare), and G3 (nighttime).
Character recognition is the last and hence the very V.4 FIELD TESTING DATASET
important step, as the evaluation of performance is based on this
step. Most of the character recognition techniques use variants of The field-testing dataset included 12000 photos from the
CNN [3]. Some use a single stage License Plate Recognition Transport Bureau of the Macao S.A.R. (DSAT) [1]. There are three
approach that uses end-to-end CNN architecture while some multi- main types of vehicles in the dataset: automobiles, trucks, and
stage approaches use a CNN architecture in their License Plate buses. The resolution of these images is 1024 x 800. Between the
recognition stage [1],[4],[11],[12]. In the papers [1-4], a variant of training and validation sets, these photos are split in a 4:1 ratio.
CNN, VGG & VGG-16 are used. In the paper [1], a multi-stage
license plate recognition is used which uses VGG to recognize the V.5 LPST-110K
license plate. VGG-16 is the modified version of VGG that is used
in [4] with some modification in the architecture to acquire better The LPST-110K [2] Dataset consists of pictures shot in
results. open spaces. It is the first dataset to simultaneously handle LP and
A 2-layer LSTMs is used for recognition of license plates scene text for LP detection. The LPST-110K is the first dataset that
[8], which eliminates the need for individual character provides text annotations in addition to a significant number of
segmentation and hence gives better performance without examples (LP and non-LP) in a picture, even when those instances
character segmentation. Resampling and rectification of LP is done are taken from scenes without any limitations. The LPST-110K
according to the vertices obtained from VertexNet. The corrected dataset compiles images from hundreds of dash cameras and
LP picture is then sent to SCR-Net. Through a forward pass, SCR- security cameras installed in moving cars and structures,
Net guesses the characters [6]. The number-plate's alphanumeric encompassing locations in East Asia and Europe. Along with the
characters are recognized using the Tesseract OCR engine. Prior LP Road signs, wallpaper text, banners, and commercial adverts
training is done to increase the Tesseract OCR engine's accuracy are also included in the collection in [2] as non-LP scene texts.
[9]. There are 9,795 photos and 110,000 scene text pieces in the LPST-
110K collection.
V. DATASET ANALYSIS The scene texts contain 51,031 LP instances and 58,969
non-LP instances. The resolution of each image in the collection is
V.1 HZM MULTI-STYLE DATASET 1280 (Width) x 720 (Height) x 3 (Channels). The photos in LPST-
The "HZM multi-style dataset" in [1] is a collection of 110K are compressed using the h264 codec setting in contrast to
automobile pictures taken from the Hong Kong-Zhuhai-Macao most other LP detection datasets.
Bridge and is referred to as such because it contains several forms
of license plates.176 photos are utilized for testing, and 1200
images are used to train the model in this proprietary dataset, which

Page 132
 One, Two and Three, ITEGAM-JETIA, Manaus, v.10 n.48, p. 129-134, July/August., 2024.

V.6 VALID CCPD-Base and the 80k examples of sub-datasets such as CCPD-
DB, CCPD-FN, CCPD-Rotate, CCPD-Tilt, CCPD-Weather, and
The two auto-mobile data recorders are used to record CCPD-Challenge.
videos in 720 x 1280 resolution on the streets of a Chinese city1.
The collected dataset is known as the "Vehicle and License Plate VI. PERFORMANCE EVALUATION APPROACHES
Dataset" (VALID [4]). A dataset includes a total 887 well
annotated images. The test set consists of 78 photos from a single Most of the ALPR system uses loss function as the
recorder. 809 additional photos from another recorder are divided evaluation method [1],[2]. This loss function is calculated at each
randomly in the ratio 7:3 into the training set and the validation set. stage in the single stage [4],[8],[9] as well as the multi-stage
approach [1-3], [5],[6]. Finally, the results are aggregated to get the
V.7 DETROIT right accuracy of overall ALPR systems. The accuracy is calculated
as per loss value at each stage [12].
The "Car" and "Vehicle registration plate" are part of the re- In the [2] paper the evaluation techniques used are
annotated DETROIT Dataset, which is a subset of the Open Image Precision, Recall, F-measure and IoU.
Dataset (OID).in a simple way DETROIT is called as (Dataset from
Open Image Dataset). DETROIT Dataset is a re-annotated subset VI.1 PRECISION
of the Open Image Dataset (OID), which contains “Car” and
“Vehicle registration plate”. For simplicity, DETROIT is called as The ratio of the number of successfully identified
(Dataset from Open Image Dataset) [4]. The size and aspect ratio bounding boxes to all acquired bounding box candidates is known
of the DETROIT photos, which are downloaded from the Internet, as precision. [2],[5-7].
can vary greatly. The test set consists of 386 images taken from the 𝑇𝑝
OID validation set. 1113 OID test photos are randomly split into a 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (1)
𝑇𝑝+𝐹𝑝
training set and a validation set in the ratio of 7:3.
Where, Tp = correctly estimated bounding box
V.8 DOC Fp = incorrectly estimated bounding box
To obtain DOC (Dataset from Cars), the location of the VI.2 RECALL
vehicle and the location of the license plate are combined. There
are overall 105 photos in the dataset. Out of which 70% chosen at The recall is the ratio of the correctly estimated bounding
random as the training-validation set while the remaining 30% are boxes among all the ground truths [2],[5],[7].
used as the test set. The size and aspect ratio of the DOC images, 𝑇𝑝
which are downloaded from the Internet, vary widely. 𝑅𝑒𝑐𝑎𝑙𝑙 = (2)
𝑇𝑝+𝐹𝑛

V.9 CLPD Where, Fn = the quantity of the undetected ground truth

There are a total of 1200 images in the CLPD [8] (China VI.3 F-MEASURE
License Plate Dataset) dataset, which comes from all 31 provinces
on the mainland. It covers a wide range of photographic situations, Benchmark for LP detection evaluation used in PKU dataset
vehicle types and regional codes, allowing for an in-depth analysis [2],[7].
(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙)
of current license plate recognition techniques while promoting the 𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = 2 ∗ (3)
(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙)
development of a more useful model. Licence plate images in the
CLPD dataset are gathered from various kinds of real-scene image VI.4 IoU
sources, such as web searches, images taken from smartphones,
and driving recorder recordings of automobiles. The photography When the detected bounding box's IoU overlaps the ground
angles, shooting times, resolutions, and background are also taken truth region by more than 50% (IoU > 0.5), it is deemed to be
into consideration when capturing LP images to account for the accurate [2],[3].
various conditions. Various vehicle types, including cars, trucks, 𝑎𝑟𝑒𝑎(𝑅𝑑𝑒𝑡∪ 𝑅𝑔𝑡)
police cars and new energy vehicles, are included in the CLPD 𝐼𝑜𝑈 = (4)
𝑎𝑟𝑒𝑎(𝑅𝑑𝑒𝑡∩ 𝑅𝑔𝑡)
dataset. The real-world dataset CLPD [6] contains a wide range of
vehicle types, environment, and area codes. Where, Rdet = area of the detected bounding box
Rgt = ground truth
V.10 CCPD AP (Average Precision) is another method used for the
evaluation of ALPR systems [2],[4]. AP is calculated over IoU
The Chinese City Parking Dataset (CCPD) [6] provides a (Intersection over Union) [2].
large-scale and comprehensive Licence Plate benchmark to The performance evaluation measure Success Ratio is
evaluate Automated License Plate Recognition techniques under used in [13]. Success Ratio is the ratio of the number of success
uncontrolled conditions. CCPD contains 280k vehicle images, samples to the total number of samples.
which is two orders of magnitude greater than other LP datasets,
𝑁𝑆𝑠
that were taken under uncontrolled conditions, such as diverse 𝑆𝑅 = ∗ 100 (5)
𝑇𝑁𝑠
weathers, lighting, rotation, and vagueness. Each image has a 720
x 1160 resolution. The dataset provides sufficient annotations, Where, SR = Success Ratio
including the LP character, bounding box, four vertices, degree of NSs = Number of success samples
tilt in both the horizontal and vertical axes, brightness, and TNs = Total number of samples
vagueness levels. The model is trained by using 100k examples of The evaluation of classification accuracy in [9] is done
CCPD-Base, and it is tested on the remaining 100k examples of using the formula below.

Page 133
 One, Two and Three, ITEGAM-JETIA, Manaus, v.10 n.48, p. 129-134, July/August., 2024.

𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = Approval of the final text: Rohini B. Kokare.

𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑐ℎ𝑎𝑟𝑎𝑐𝑡𝑒𝑟𝑠
(6)
𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑐ℎ𝑎𝑟𝑎𝑐𝑡𝑒𝑟 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
VIII. REFERENCES
As seen in most references the performance measures are [1] Qiuying Huang , Zhanchuan Cai and Ting La, A New Approach for Character
accuracy, precision, recall, F-measure, AP/IoU. In most multi- Recognition of Multi-Style Vehicle License Plates, IEEE TRANSACTIONS ON
MULTIMEDIA, VOL. 23, 2021
stage approaches the loss function is the best estimation of https://doi.org/10.1109/TMM.2020.3031074
accuracy.
[2]Younkwan Lee, Jihyo Jeon, Yeongmin Ko, Moongu Jeon and Witold Pedrycz,
VII. OPEN CHALLENGES FOR FUTURE License Plate Detection via Information Maximization, IEEE TRANSACTIONS
ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 9,
VII.1 VARIATIONS IN BACKGROUND COLOR SEPTEMBER 2022.
https://doi.org/10.1109/TITS.2021.3135015
License plates do have different background colors and each
different color signifies different types of the vehicle. White [3] Chunsheng Liu and Faliang Chang Hybrid Cascade Structure for License Plate
background is for personal vehicles while the yellow one is for Detection in Large Visual Surveillance Scenes, IEEE TRANSACTIONS ON
INTELLIGENT TRANSPORTATION SYSTEMS, Volume: 20, Issue: 6, June
transport vehicles and so on. These different colors can add 2019.
complexity in terms of pre-processing [3]. This is a prominent https://doi.org/10.1109/TITS.2018.2859348
future scope to work upon.
[4] Song-Lu Chen , Chun Yang, Jia -Wei Ma, Feng Chen, and Xu-Cheng Yin ,
Simultaneous End-to-End Vehicle and License Plate Detection With Multi-Branch
VII.2 SPEED OF TRAINING Attention Neural Network, IEEE TRANSACTIONS ON INTELLIGENT
TRANSPORTATION SYSTEMS, Volume: 21, Issue: 9, September 2020.
Image processing requires a significant amount of time. To https://doi.org/10.1109/TITS.2019.2931791
fasten the time decoders are used. Some decoders perform
sequentially and not parallelly. LSTM is such a decoder that is [5]Xiangjie Kong, Kailai Wang, Mingliang Hou, Xinyu Hao, Guojiang
incapable of parallel training [8]. We can try out using a Shen, Xin Chen, and Feng Xia, A Federated Learning-based License Plate
Recognition Scheme for 5G-enabled Internet of Vehicles, IEEE TRANSACTI ONS
transformer like decoder to make training faster. ON INDUSTRIAL INFORMATICS, Volume:17, Issue:12, December 2021
https://doi.org/10.1109/TII.2021.3067324
VII.3 CHALLENGING WEATHER CONDITIONS [
[6]Yi Wang, Zhen-Peng Bian, Yunhao Zhou, and Lap-Pui Chau, Rethinking and
Changing weather conditions affect the lumination and Designing a High-Performing Automatic License Plate Recognition Approach,
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEM S,
hence the performance of ALPR systems [7]. To adapt with the 2021.
weather conditions is a major future issue. More challenging is to https://doi.org/10.1109/TITS.2021.3087158
detect LPs at nighttime, as the luminance is lowest at that time.
[7]Meeras Salman Al-Shemarry , Yan Li , and Shahab Abdulla, An Efficient
Texture Descriptor for the Detection of License Plates From Vehicle Images in
VIII. CONCLUSIONS Difficult Conditions, IEEE TRANSACTIONS ON INTELLIGENT
TRANSPORTATION SYSTEMS, Volume: 21, Issue: 2, February 2020
In account for various operational and hardware constraints, https://doi.org/10.1109/TITS.2019.2897990
a careful selection of standards and techniques is required for the
design and development of an automatic license plate recognition [8]Linjiang Zhang, Peng Wang, Hui Li , Zhen Li, Chunhua Shen, and
system (ALPR). This research article has investigated and Yanning Zhang, A Robust Attentional Framework for License Plate
Recognition in the Wild, IEEE TRANSACTIONS ON INTELLIGENT
examined the currently employed ways and strategies in the most TRANSPORTATION SYSTEMS, Volume: 22, Issue: 11, November 2021.
recent literature on ALPR solutions. The deep single-stage https://doi.org/10.1109/TITS.2020.3000072
Learning-based systems have demonstrated strong results with
various datasets. Although learning systems can be pre-trained on [9] Sparsh Jain, Rishikesh Rathi, and Rahul Kumar Chaurasiya, Indian Vehicle
Number-Plate Recognition using Single Shot Detection and OCR, IEEE INDIAN
huge datasets, single-stage approaches have demonstrated greater COUNCIL INTERNATIONAL SUBSECTIONS CONFERENCE, 2021.
computing efficiency and accuracy. This survey did a job of https://doi.org/10.1109/INDISCON53343.2021.9582216
thorough analysis of linked studies and identified the specifications
for actual benchmark datasets. We have outlined the current issues [10] Zijia Liang, Erkang Li, Yanxing Huo, Yutong Guo, Chengning Lu, Design of
Portable License Plate Recognition System, 2021 IEEE INTERNATIONAL
and made future directions for ALPR study. CONFERENCE ON POWER, INTELLIGENT COMPUTING AND SYSTEM S
(ICPICS), 2021.
VI. AUTHOR’S CONTRIBUTION https://doi.org/10.1109/ICPICS52425.2021.9524092

Conceptualization: Vishakha H. Jagtap, Rohit V. Dhotre, Utkarsh [11] Chenxu Duan, Shiqiang Luo, Design of License Plate Recognition System
R. Khandare, Harshada N. Khuspe, Rohini B. Kokare. Based on OpenCV, 2022 15TH INTERNATIONAL SYMPOSIUM ON
COMPUTATIONAL INTELLIGENCE AND DESIGN (ISCID), 2022.
Methodology: Vishakha H. Jagtap, Rohit V. Dhotre.
https://doi.org/10.1109/ISCID56505.2022.00031
Investigation: Vishakha H. Jagtap, Rohit V. Dhotre, Utkarsh R.
Khandare, Harshada N. Khuspe. [12] Liang Wang, Yimei Huanga, Chengqun Lianga, Jinrong Zhoua, Taoqiang
Discussion of results: Vishakha H. Jagtap, Rohit V. Dhotre, Zhua, License plate recognition system based on image recognition, 2022 IEEE
21ST INTERNATIONAL CONFERENCE ON UBIQUITOUS COMPUTING
Utkarsh R. Khandare, Harshada N. Khuspe, Rohini B. Kokare. AND COMMUNICATIONS (IUCC/CIT/DSCI/SMARTCNS), 2022.
Writing – Original Draft: Vishakha H. Jagtap, Rohit V. Dhotre,
Utkarsh R. Khandare, Harshada N. Khuspe. 13] Milan Samantaray, Anil Kumar Biswal, Debabrata Singh, Debabrata Samanta,
Writing – Review and Editing: Vishakha H. Jagtap. Marimuthu Karuppiah, Niju P Joseph, Optical Character Recognition (OCR) based
Vehicle’s License Plate Recognition System Using Python and OpenCV, 2021 5TH
Resources: Rohit V. Dhotre, Utkarsh R. Khandare, Harshada N. INTERNATIONAL CONFERENCE ON ELECTRONICS, COMMUNICATI ON
Khuspe. AND AEROSPACE TECHNOLOGY (ICECA), 2021.
Supervision: Rohini B. Kokare. https://doi.org/10.1109/ICECA52323.2021.9676015

Page 134