(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 15, No. 1, 2024

A Deep Learning-based Framework for Vehicle

License Plate Detection
Deming Yang1*, Ling Yang2
Huanghai College of Automotive Engineering, Liaoning Jidian Polytechnic, Dandong 118009, Liaoning, China
Department of Information Engineering, Liaoning Jidian Polytechnic, Dandong 118009, Liaoning, China

Abstract—In the contemporary landscape of smart cameras, radar systems, as well as connected vehicles
transportation systems, the imperative role of intelligent traffic contribute to the data collection process [9]. Meanwhile,
monitoring in bolstering efficiency, safety, and sustainability advanced algorithms and analytics techniques, including deep
cannot be overstated. Leveraging recent strides in computer learning [10], have emerged as powerful tools for traffic
vision, machine learning, and data analytics, this study addresses analysis, pattern recognition, and prediction.
the pressing need for advancements in car license plate
recognition within these systems. Employing an innovative The research significance of intelligent traffic monitoring in
approach based on the YOLOv5 architecture in deep learning, smart transportation is paramount. Accurate and efficient
the study focuses on refining the accuracy of license plate monitoring systems are crucial for traffic management
recognition. A bespoke dataset is meticulously curated to authorities, urban planners, and policymakers to make
facilitate a comprehensive evaluation of the proposed informed decisions regarding infrastructure planning, traffic
methodology, with extensive experiments conducted and metrics control strategies, and resource allocation. Furthermore,
such as precision, recall, and F1-score employed for assessment. intelligent traffic monitoring has the potential to improve safety
The outcomes underscore the efficacy of the approach in by enabling early detection of traffic incidents and facilitating
significantly enhancing the precision and accuracy of license timely emergency responses. It also contributes to reducing
plate recognition using performance evaluation of the proposed energy consumption, mitigating environmental impacts, and
method. This tailored dataset ensures a rigorous evaluation,
enhancing overall transportation system resilience.
affirming the practical viability of the proposed approach in real-
world scenarios. The study not only showcases the successful Among the various technologies used in intelligent traffic
application of deep learning and YOLOv5 in achieving accurate monitoring [11], vision-based systems have attracted
license plate detection and recognition but also contributes to the considerable attention from researchers [12]. Vision-based
broader discourse on advancing intelligent traffic monitoring for systems use computer vision techniques to extract relevant
more robust and efficient smart transportation systems. information from visual data, like images or videos captured by
surveillance cameras. The appeal of vision-based approaches
Keywords—Intelligent traffic monitoring; smart transportation;
lies in their ability to provide rich and detailed information
deep learning; Yolov5; performance evaluation
about traffic conditions, including vehicle movements, license
I. INTRODUCTION plate recognition [13], and traffic flow analysis. These systems
have been continuously evolving, with the introduction of more
Intelligent Traffic Monitoring plays a vital role in the sophisticated computer vision algorithms and deep learning-
development of smart transportation systems, aiming to based methods.
enhance the efficiency, safety, and sustainability of modern
urban mobility [1], [2]. By utilizing advanced technologies and Several studies have focused on computer vision-based
data-driven approaches, intelligent traffic monitoring enables intelligent traffic monitoring systems. These studies have
real-time analysis, prediction, and management of traffic explored different aspects, such as object detection, tracking,
conditions [3], [4]. It encompasses various components such as license plate recognition, and traffic flow estimation [14], [15].
sensor networks, data integration, traffic analysis algorithms, Fig. 2 shows the schematic of a license plate recognition
and intelligent decision-making systems [5]. Efficient traffic system. Recently, deep learning-based approaches [16],
monitoring is essential for optimizing traffic flow, reducing particularly those employing You Only Look Once (YOLO)
congestion, and improving overall transportation experiences. architecture [17], have gained attention due to their superior
performance in object detection tasks [18], [19,30]. These
Recent years have witnessed significant advancements in methods leverage large-scale datasets and powerful deep neural
intelligent traffic monitoring technologies, driven by the rapid network architectures to achieve high accuracy as well as real-
progress in computer vision, machine learning, and data time processing capabilities [20]. However, despite these
analytics[6], [7]. Fig. 1 demonstrates the schematic of an advancements, there still exist certain limitations and research
intelligent transportation system [8]. These advancements have gaps that need to be addressed to enhance further the
enabled the development of sophisticated systems that can effectiveness and robustness of intelligent traffic monitoring
automatically capture, process, and analyze vast amounts of systems.
traffic data in real-time. Technologies like video surveillance

1009 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

Fig. 1. Schematic of an intelligent transportation system.

Fig. 2. Schematic of a license plate recognition system [21].

This study is motivated by the essential role of Intelligent plate recognition. The aim is to contribute to the field by
Traffic Monitoring in advancing smart transportation systems conducting extensive experiments with a custom dataset,
for enhanced efficiency, safety, and sustainability in urban encompassing training, validation, and testing processes, to
mobility. Recent technological advancements, particularly in enhance the effectiveness and robustness of intelligent traffic
computer vision and machine learning, have enabled monitoring systems.
sophisticated monitoring systems, but existing approaches still
have limitations. The focus is on vision-based systems, This study presents a method that uses a deep learning
utilizing computer vision techniques for detailed traffic framework with the YOLOv5 architecture to improve car
information, including license plate recognition. To address license plate recognition. To evaluate the effectiveness of the
research gaps, the study proposes a novel approach using deep proposed approach, generating a custom dataset and conduct
learning with the YOLOv5 architecture to improve car license extensive experiments involving training, validation, and
testing processes.

1010 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

The following is a list of the major research contributions classifier for character recognition. The system achieves high
made in this article: accuracy in license plate recognition, even in challenging
scenarios with variations in lighting conditions, plate
1) Improving the accuracy of car license plate recognition orientations, and background clutter. The findings show that
by proposing a novel approach based on a deep learning the suggested system outperforms existing methods on
technique with the YOLOv5 architecture, resulting in superior benchmark datasets, demonstrating its robustness and
recognition performance compared to existing methods. effectiveness. However, the limitations of the system lie in its
2) Developing and utilizing a custom dataset sensitivity to complex backgrounds and the requirement for
encompassing diverse car license plate images to facilitate well-segmented license plate regions.
comprehensive evaluation and benchmarking of the proposed Kaur et al. [25] presented an automatic license plate
YOLOv5-based approach for license plate recognition. recognition system for vehicles using a convolutional neural
3) Conducting extensive experiments to assess the network (CNN). The method involves training the CNN on a
effectiveness as well as the performance of the suggested large dataset of license plate images to learn the patterns and
YOLOv5-based approach, including training, validation, and features necessary for accurate recognition. Key features of the
testing processes, and evaluating metrics regarding F1-score, system include the ability to extract license plate regions,
recall and precision performance metrics. segment individual characters, and recognize them using the
trained CNN. The findings illustrate that the suggested CNN-
The remainder of this article includes the following parts: based approach achieves high accuracy in license plate
Section I presents the introduction. The literature review is recognition, outperforming traditional methods. However, the
discussed in Section II. Section III presents the material and system's limitation lies in its dependence on well-labeled
method. Experimental results are discussed in Section IV. training data and its susceptibility to variations in license plate
Finally, Section V concludes the paper. designs, environmental conditions, and image quality.
II. RELATED WORKS A low-cost Internet of Things (IoT) based Arabic license
Wang et al. [22] proposed a light convolutional neural plate recognition model for smart parking systems presented by
network (CNN) approach for end-to-end car license plate Abdellatif et al. [26]. The method uses a combination of
recognition and detection. The method involves training a machine learning algorithms and image processing techniques
lightweight CNN model that can directly detect and recognize to detect as well as recognize Arabic license plates. Key
license plates in images. The key features of the approach features of the approach include its cost-effectiveness, reliance
include its simplicity, efficiency, and ability to handle various on IoT infrastructure, as well as the ability to recognize Arabic
license plate designs and environmental conditions. The license plates accurately. The findings demonstrate the
findings show that the proposed light CNN achieves effectiveness of the suggested model in real-world scenarios,
competitive performance regarding accuracy and processing achieving high accuracy in license plate recognition for smart
speed, outperforming traditional methods. However, the parking systems. However, the limitation of the system lies in
limitation of the method lies in its dependence on well-labeled its focus on Arabic license plates, limiting its applicability to
training data and potential challenges in handling extremely other regions with different license plate formats. Further
blurred or distorted license plate images. Further research is research is recommended to explore the model's
required to address these limitations and enhance the system's generalizability to different languages and license plate
adaptability to complex real-world scenarios. designs.

Pustokhina et al. [23] developed an automatic vehicle Shi et al. [27] proposed a License Plate Recognition System
license plate recognition system for intelligent transportation (LPRS) that combines an improved version of the YOLOv5
systems, utilizing optimal K-means clustering in combination object detection algorithm and a GRU model for accurate
with a convolutional neural network (CNN). The method license plate recognition and detection. The method involves
demonstrates high accuracy and efficiency in recognizing two stages: license plate detection using the improved
license plate characters, leveraging K-means clustering for YOLOv5, which incorporates feature fusion and attention
region extraction and a CNN for character segmentation and mechanism to enhance the detection performance, and license
recognition. The system achieves competitive performance plate recognition using a GRU model, which learns the
compared to existing approaches and proves effective in sequential patterns of characters on the license plate. The key
diverse scenarios. However, the paper acknowledges features of the proposed system include improved accuracy in
limitations related to region extraction accuracy and license plate recognition and detection, robustness to varying
adaptability to varying license plate designs and environmental lighting and environmental conditions, and efficient processing
conditions, suggesting further research for enhancing system speed. The experimental outcomes demonstrate superior
robustness. performance compared to existing methods, achieving high
accuracy rates in license plate recognition and detection tasks.
He et al. [24] presented a robust automatic recognition However, the limitation of the system lies in its reliance on a
system for Chinese license plates in natural scenes. The large dataset for training, which may pose challenges in
proposed method employs a combination of image scenarios with limited data availability.
preprocessing, character segmentation, and recognition
algorithms. The key features include adaptive thresholding, Despite the notable advancements in vehicle license plate
connected component analysis, as well as a CNN-based recognition (VLPR) systems, a research gap persists in

1011 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

achieving a more accurate and robust method for license plate addressing real-world challenges and demonstrating potential
detection. Although existing studies, have made significant advancements in the field.
contributions by introducing various approaches combining the
CNNs, clustering techniques, and innovative algorithms for The following categories apply to the steps involved in
license plate detection and recognition. However, common model generation in YOLOv5:
limitations across these studies include dependencies on well- 1) Data preparation: Gather a labeled dataset of license
labeled training data, challenges in handling blurred or plate images. Annotate the license plate regions with bounding
distorted license plate images, adaptability issues to varying boxes and assign corresponding labels.
designs and environmental conditions, and sensitivity to
2) Model training: Train the YOLOv5 model on the
complex backgrounds. While these studies demonstrate
promising results, the need for a more accurate and adaptable annotated license plate dataset. This involves optimizing the
VLPR system that addresses these limitations is evident. model's parameters to detect license plates accurately.
Therefore, the research gap lies in developing a method that 3) Model evaluation: Utilize evaluation metrics like mean
not only surpasses current accuracy rates but also ensures average precision (mAP) and detection accuracy to assess the
robustness in challenging real-world scenarios, such as performance of the trained model. This step helps determine
scenarios with limited training data availability or diverse the model's effectiveness in license plate recognition.
license plate designs and environmental conditions. Closing 4) Inference: Utilize the trained YOLOv5 model to
this gap is crucial for advancing the field and facilitating the perform license plate detection on unseen images or videos.
deployment of more reliable and effective vehicle license plate The model identifies the license plate regions and provides the
detection systems in smart transportation and intelligent traffic
corresponding bounding boxes and class predictions.
monitoring applications.
5) Post-processing: Apply post-processing methods like
III. METHODOLOGY non-maximum suppression (NMS) to filter out redundant
In this study, the researchers propose an algorithm for bounding boxes and retain the most accurate license plate
license plate recognition that is based on the YOLOv5 model. detections.
The algorithm leverages the strengths of YOLOv5, which is B. Data Preparation
known for its real-time object detection capabilities, to
recognize license plates in images or videos effectively. By Preparing data for training a YOLOv5 model using
using YOLOv5 inspired by [17], [28], [29], the algorithm aims collected images involves two key steps: image annotation and
to achieve accurate as well as efficient license plate data augmentation. This study collected images from internet
recognition, offering potential applications in various domains, resources. In image annotation, the specific objects of interest,
such as traffic monitoring for smart transportation applications. such as license plates, are manually labeled by drawing
bounding boxes around them and assigning corresponding
In this study, a method for license plate recognition is class labels. This step ensures that the model learns to detect
proposed, capitalizing on the strengths of the YOLOv5 model the desired objects accurately. Data augmentation, on the other
renowned for its real-time object detection capabilities. The hand, involves applying various transformations to the images
algorithm aims to achieve both accuracy and efficiency in to expand the training dataset artificially. Techniques like
license plate recognition, with potential applications in diverse cropping, rotation, scaling, flipping, color jittering, and noise
domains, particularly in smart transportation for traffic injection are used to simulate real-world variations, improving
monitoring. The proposed methodology involves a multi-step the model's ability to generalize to various scenarios.
process for model generation within the YOLOv5 framework.
The first step, Data Preparation, entails collecting a labeled By combining image annotation and data augmentation, a
dataset of license plate images, annotating license plate custom dataset is created for training a YOLOv5 model. The
regions, and assigning corresponding labels. Model Training annotated images provide the necessary ground truth
follows, where the YOLOv5 model is trained on the annotated information, enabling the model to learn the spatial location
dataset, optimizing parameters for accurate license plate and class labels of license plates. Data augmentation diversifies
detection. The subsequent step, Model Evaluation, utilizes the dataset, introducing variations in object appearance, scale,
metrics like mean average precision (mAP) and detection and environmental conditions. This helps the model become
accuracy to assess the trained model's performance. Inference more robust and generalize well to unseen license plate images.
involves utilizing the trained YOLOv5 model for license plate These steps prepare the data as well as ensure the effectiveness
detection on unseen images or videos. The final step, post- and accuracy of the YOLOv5 model for license plate detection
processing, employs techniques such as non-maximum tasks.
suppression (NMS) to refine predictions and ensure the most C. Model Training
confident and accurate license plate detections are retained.
To generate a YOLOv5 model using a custom dataset and
Noteworthy efforts are invested in Data Preparation, which
split it into training, testing, and validation sets, you can follow
includes image annotation and data augmentation for creating a
these steps:
custom dataset that enhances the model's robustness and
generalization capabilities. Model Training involves splitting 1) Dataset split: First, split the custom dataset into three
the dataset, training the YOLO model, and fine-tuning based subsets: testing, validation, also training. Allocate 75% of the
on the validation set. The proposed algorithm provides a data for training, 15% for validation, also the remaining 10%
comprehensive approach to license plate recognition,

1012 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

for testing. Ensure that the splits are representative and 2) Precision: Precision measures the proportion of
maintain a balanced distribution of license plate images across correctly detected license plates among all the predicted
all subsets. license plates. It is computed as the ratio of true positives
2) Training module: The training module involves training (correct detections) to the sum of false positives as well as true
the YOLO model on the training dataset. During training, the positives (incorrect detections).
model learns to detect license plates by optimizing its 3) Recall: Recall measures the proportion of correctly
parameters utilizing a training algorithm like Adam optimizer detected license plates among all the ground truth license
or stochastic gradient descent (SGD). The training process plates. It is computed as the ratio of true positives to the total
iteratively updates the model's weights based on the loss of false negatives also true positives (missed detections).
function, which consists of localization loss (IoU loss) as well 4) F1-score: The F1-score, which provides a balanced
as confidence loss (objectness loss). The model is exposed to assessment of the model's performance, is the harmonic mean
the training dataset, and backpropagation is used to update the of precision and recall. It is computed as 2 * ((precision *
weights, gradually improving the model's accuracy in recall) / (precision + recall)).
detecting license plates.
Examine the recall, precision, and F1-score values to assess
3) Validation module: The validation module is used to the model's performance in license plate recognition. Greater
assess the model's performance and tune its hyperparameters precision signifies fewer false positive detections, whereas
during training. The validation set, comprising 15% of the higher recall signifies fewer false negatives. The F1-score
custom dataset, is utilized to evaluate the model's accuracy provides an overall evaluation by considering both precision
and generalization ability. The trained model is run on the and recall. The details of performance metrics results are
validation set, and metrics like mean average precision (mAP) presented in the following section.
are calculated to measure the quality of license plate detection. E. Inference
To optimize the model's performance, adjustments to the
Inference refers to the process of utilizing a trained
training procedure, model architecture, or hyperparameters YOLOv5 model to perform license plate detection on new,
might be made in light of the validation results. unseen images or videos. Once the YOLOv5 model has been
4) Testing module: The testing module is the final stage, trained on a dataset, it learns to recognize license plates by
where the trained YOLO model is appraised on the testing detecting and localizing them within an image. During
dataset, which also consists of 10% of the custom dataset. The inference, the trained model takes an input image or frame
testing dataset contains unseen license plate images that were from a video and processes it through the model's architecture.
not used during training or validation. The model's The inference process involves passing the input image
performance is assessed by running it on the testing dataset, through the YOLOv5 model, which applies a series of
also measuring metrics like mAP, detection accuracy, and convolutional layers, down-sampling, and feature extraction to
false positives/negatives. This module provides an identify potential license plate regions. The model predicts
understanding of the model's real-world performance as well bounding box coordinates as well as class probabilities for each
as its ability to recognize license plates in unseen and new detected object, including license plates. These predictions are
scenarios accurately. generated based on the learned patterns and features acquired
during the training phase. After running inference, the
D. Model Evaluation YOLOv5 model provides the predicted bounding box
Assessing the model is vital to validate the proficiency of a coordinates and class labels for detected license plates. This
YOLOv5 model for recognizing license plates. It provides a information allows for precise localization of license plates
means to objectively assess the model's performance, measure within the image or video frame. The model's output can be
its accuracy and generalization ability, and identify areas for further processed to extract the license plate region and
improvement. By evaluating the model on independent perform subsequent tasks such as character recognition or
datasets, stakeholders can gain confidence in its effectiveness, vehicle identification.
ensure it meets desired requirements, and make informed
F. Post-processing
decisions regarding its deployment and usage. To evaluate a
generated YOLOv5 model for license plate recognition using In the post-processing stage of license plate recognition
performance metrics like F1-score, recall, and precision, the using a trained YOLOv5 model, techniques such as non-
following steps are: maximum suppression (NMS) are applied to refine the model's
predictions. NMS helps eliminate redundant bounding boxes
1) Intersection over Union (IoU) calculation: For each by retaining only the most confident and accurate detections.
predicted bounding box, calculate the Intersection over Union By comparing the confidence scores and utilizing intersection
(IoU) with the corresponding ground truth bounding box. IoU over union (IoU) calculations, NMS suppresses overlapping
examines the overlap between the predicted and ground truth detections and ensures that only the highest-scoring detection
bounding boxes to determine if detection is a true positive or a for each object is retained. This post-processing step improves
false positive. the quality and accuracy of license plate recognition results by
removing duplicate detections and producing cleaner outputs.

1013 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

IV. RESULTS AND PERFORMANCE EVALUATION Based on these definitions, the ability of a model to
Performance evaluation and experimental outcomes for a accurately identify positive predictions is measured by its
YOLOv5 model in license plate recognition involve assessing precision. It is calculated using the following formula:
the model's performance utilizing metrics like accuracy, recall, 
precision, F1-score and mAP. Fig. 3 shows a sample result of
the license plate recognition. The model is tested on a separate Furthermore, the recall measures the proportion of correctly
dataset with ground truth annotations, and its predictions are predicted license plates out of all actual license plates. It is
compared against these annotations to calculate the metrics. calculated using the following formula:
The results provide quantitative measures of the model's ability 
to detect and localize license plates accurately.
Precision quantifies the proportion of correctly predicted
The analysis of experimental results allows for an license plates out of all predicted license plates. A higher
evaluation of the YOLOv5 model's overall performance and precision demonstrates fewer false positives, meaning that the
helps identify areas for improvement. By examining metrics model is more accurate in identifying true license plates. The
like precision, accuracy, recall, and F1-score, the model's recall quantifies the model's ability to recognize all positive
effectiveness in license plate recognition can be assessed. The instances correctly. A higher recall demonstrates fewer false
mAP metric provides insights into the trade-off between negatives, indicating that the model can successfully detect
precision and recall, while IoU analysis helps gauge the most of the license plates present. Fig. 4 and Fig. 5 show the
localization accuracy. Comparisons with other models or result of precision and recall performance metrics.
benchmarks further contribute to understanding the model's
relative strengths and weaknesses. These findings guide As displayed in Fig. 4, precision is a performance metric
iterative improvement, enabling researchers to refine the utilized to appraise the accuracy of a YOLOv5 model in
YOLOv5 model's architecture, training process, or other license plate recognition. It measures the proportion of
parameters to enhance its license plate recognition capabilities. correctly detected license plates among the predicted ones. In
precision evaluation, the Y-axis demonstrates precision values,
For the license plate recognition using the YOLOv5 model, while the X-axis represents confidence values or thresholds.
performance metrics such as precision are commonly The precision curve visualizes the trade-off between precision
employed to assess the model's accuracy. For this evaluation, as well as recall at various confidence thresholds, helping
following performance metrics are used, determine the optimal threshold that balances precision and
True Positive (TP): It refers to the cases where the recall. Higher precision indicates fewer false positives, making
YOLOV5 model correctly predicts a license plate as positive it important for tasks like license plate recognition in real-
(i.e., correctly identifying a license plate when one exists). world scenarios.
True Negative (TN): It represents the cases where the As shown in Fig. 5, in the recall evaluation, the Y-axis
YOLOV5 model correctly predicts the absence of a license denotes the recall values, while the X-axis denotes the
plate as negative (i.e., correctly identifying the absence of a confidence values or the threshold applied to the model's
license plate when none exists). predictions. To calculate recall, the forecasted bounding boxes,
as well as their associated confidence scores, are sorted in
False Positive (FP): It occurs when the YOLOV5 model descending order. A threshold is set on the confidence scores to
incorrectly predicts a license plate when there isn't one (i.e., determine which detections are regarded as positive forecasts.
incorrectly identifying a license plate when none exists). By varying the threshold, different levels of confidence are
False Negative (FN): It refers to the cases where the required for a detection to be considered positive. The recall is
YOLOV5 model fails to predict a license plate when one exists then computed as the ratio of true positives (correctly detected
(i.e., incorrectly not identifying a license plate when one license plates) to the total of true positives also false negatives
exists). (missed detections).

Fig. 3. Samples result of the license plate recognition.

1014 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

Fig. 4. The result of precision metric.

Fig. 5. Result of recall metric.

Moreover, as shown in Fig. 5, the recall curve is a graphical into the model's ability to detect license plates across different
representation of recall as a function of the confidence confidence thresholds.
threshold. The curve is generated by plotting the recall values
at different confidence thresholds on the Y-axis against the The F1-score is the harmonic mean of a recall and precision
corresponding confidence values on the X-axis. As the and is computed utilizing the following formula:
threshold decreases (lower confidence threshold), more 
detections are considered positive, resulting in increased recall
but potentially lower precision. The recall curve gives insights

1015 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

The F1-score curve and how it relates to the YOLOv5 bounding box and its associated label (license plate) are
model evaluation for license plate recognition. As displayed in considered valid or not. The F1-score curve is obtained by
Fig. 6, the Y-axis of the F1-score curve represents the F1-score plotting the F1-scores at different confidence thresholds. By
itself. It is a continuous range of values between 0 and 1, where varying the threshold, the model's precision and recall trade-off
1 represents a perfect F1-score (indicating high precision and can be analyzed. Generally, as the confidence threshold
recall), and 0 represents the worst score (indicating poor increases, the model becomes more conservative in its
performance in both precision and recall). The X-axis of the predictions, resulting in higher precision but potentially lower
F1-score curve typically demonstrates the confidence values or recall. Conversely, lowering the threshold may lead to higher
the decision thresholds used by the YOLOv5 model. During recall but lower precision. The F1-score curve provides
inference, the YOLOv5 model predicts bounding boxes for insights into the model's performance at various confidence
license plates along with their confidence scores, indicating thresholds, allowing you to choose a suitable threshold based
how confident the model is in its predictions. The confidence on ythe specific requirements.
threshold is a value used to determine whether a predicted

Fig. 6. Result of F1-score metric

Fig. 7. Result of precision-recall (PR) curve

1016 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

As shown in Fig. 7, the precision-recall curve is a graphical enhancing accuracy and efficiency. The development of a
representation utilized to measure the performance of the custom dataset and the thorough evaluation through extensive
YOLOv5 model in license plate recognition tasks. The Y-axis experiments, utilizing metrics like recall, precision, and F1-
represents precision values ranging from 0 to 1, where higher score, substantiate the efficacy of the proposed methodology.
values display better accuracy in positive predictions. The X- Despite the promising outcomes, it is essential to acknowledge
axis represents recall values, also ranging from 0 to 1, where certain limitations. The dependence on well-labeled training
higher values indicate a higher ability to detect positive data and potential challenges in handling extremely blurred or
instances. The precision-recall curve for the YOLOv5 model distorted license plate images pose constraints on the
evaluation in license plate recognition showcases how adaptability of the approach to complex real-world scenarios.
precision and recall vary with changing confidence thresholds. Future research endeavors could address these limitations by
Adjusting the threshold allows for different trade-offs between exploring innovative techniques for handling diverse and
recall and precision. Rising the threshold tends to increase challenging environmental conditions. Additionally, an
precision but may result in lower recall, while decreasing the exciting avenue for further exploration lies in integrating
threshold may raise recall at the expense of precision. By multiple sensors for comprehensive data collection, potentially
analyzing the curve, users can evaluate the model's leading to the development of real-time decision-making
performance at different thresholds as well as select the systems for traffic control and optimization. This study, while
appropriate threshold based on their priorities. Whether providing a valuable contribution to the field, sets the stage for
prioritizing precision to minimize false positives or recall to future investigations aimed at advancing intelligent traffic
minimize false negatives, the precision-recall curve offers a monitoring systems and ultimately contributing to the ongoing
visual representation to assess and fine-tune the YOLOv5 evolution of smart transportation. Furthermore, this study
model's performance for license plate recognition tasks. acknowledges the limitations of the proposed approach,
particularly its reliance on well-labeled training data and
The discussion of the presented results involves a potential challenges in handling distorted license plate images,
comprehensive evaluation of the YOLOv5 model's
future research could delve into the development of more
performance in license plate recognition, utilizing various robust algorithms capable of addressing these complexities.
metrics such as accuracy, recall, precision, F1-score, and mAP. Exploring advanced image processing techniques or
The model undergoes testing on a separate dataset with ground incorporating additional pre-processing steps may enhance the
truth annotations, and the quantitative measures obtained offer adaptability of the system to varying real-world scenarios.
insights into its ability to accurately detect and localize license Moreover, there is a promising avenue for research in the
plates. The precision metric is utilized to assess the model's integration of multiple sensors, such as cameras, LIDAR, and
accuracy, considering true positives (correctly predicted license radar, to create a more comprehensive and diverse dataset for
plates), true negatives (correctly predicted absence of license training. This holistic approach could significantly contribute
plates), false positives (incorrect predictions of license plates), to the system's ability to handle different environmental
and false negatives (missed predictions). The precision curve, conditions and improve generalization across various
as illustrated in Fig. 4, demonstrates the trade-off between scenarios. Additionally, the realization of real-time decision-
precision and recall at different confidence thresholds, aiding making systems for traffic control and optimization holds great
in determining an optimal threshold for balancing accuracy. potential, necessitating research efforts to design and
Similarly, the recall metric, depicted in Figure 5, measures the implement algorithms [30] capable of providing instantaneous
model's ability to correctly predict license plates and is responses to dynamic traffic conditions. By addressing these
analyzed across various confidence thresholds. The recall research directions, future studies can contribute to the ongoing
curve offers insights into the model's performance at different evolution of intelligent traffic monitoring systems, ensuring
thresholds, showcasing the trade-off between recall and their robustness, adaptability, and efficiency in the ever-
precision. The F1-score, presented in Fig. 6, represents the changing landscape of smart transportation.
harmonic mean of precision and recall, providing a balanced
assessment of the model's performance across different ACKNOWLEDGMENT
confidence thresholds. Finally, the precision-recall (PR) curve,
as shown in Fig. 7, visually represents the model's performance This work was supported by Project of Liaoning
in license plate recognition, enabling the analysis of trade-offs Mechatronics Polytechnic College; Project number
between precision and recall at different confidence thresholds. JYLX2021029.
These results not only contribute to a comprehensive REFERENCES
understanding of the YOLOv5 model's strengths and
[1] R. Li, S. Wang, P. Jiao, and S. Lin, “Traffic control optimization
weaknesses but also guide potential refinements in its strategy based on license plate recognition data,” Journal of traffic and
architecture, training processes, and parameters to enhance transportation engineering (English edition), vol. 10, no. 1, pp. 45–57,
license plate recognition capabilities. 2023.
[2] N.-A.- Alam, M. Ahsan, M. A. Based, and J. Haider, “Intelligent system
V. CONCLUSION for vehicles number plate detection and recognition using convolutional
This study introduces a novel and improved strategy for car neural networks,” Technologies (Basel), vol. 9, no. 1, p. 9, 2021.
license plate recognition, leveraging the YOLOv5 architecture [3] V. S. R. Kosuru, A. K. Venkitaraman, V. D. Chaudhari, N. Garg, A.
Rao, and A. Deepak, “Automatic Identification of Vehicles in Traffic
within a deep learning framework. The proposed approach using Smart Cameras,” in 2022 5th International Conference on
demonstrates a commitment to achieving superior performance Contemporary Computing and Informatics (IC3I), IEEE, 2022, pp.
compared to existing methods, with a particular focus on 1009–1014.

1017 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 1, 2024

[4] N. A. Khan, N. Z. Jhanjhi, S. N. Brohi, R. S. A. Usmani, and A. Nayyar, [17] A. R. Youssef, A. A. Ali, and F. R. Sayed, “Real-time Egyptian License
“Smart traffic monitoring system using unmanned aerial vehicles Plate Detection and Recognition using YOLO,” International Journal of
(UAVs),” Comput Commun, vol. 157, pp. 434–443, 2020. Advanced Computer Science and Applications, vol. 13, no. 7, 2022.
[5] N. do V. Dalarmelina, M. A. Teixeira, and R. I. Meneguette, “A real- [18] D. J. I. Z. Chen, “Automatic vehicle license plate detection using K-
time automatic plate recognition system based on optical character means clustering algorithm and CNN,” Journal of Electrical Engineering
recognition and wireless sensor networks for ITS,” Sensors, vol. 20, no. and Automation, vol. 3, no. 1, pp. 15–23, 2021.
1, p. 55, 2019. [19] R.-C. Chen, “Automatic License Plate Recognition via sliding-window
[6] S. Nayak and K. Katakiya, “Machine Vision Based Intelligent Traffic darknet-YOLO deep learning,” Image Vis Comput, vol. 87, pp. 47–56,
Management Tool,” IJRAR-International Journal of Research and 2019.
Analytical Reviews (IJRAR), vol. 6, no. 2, pp. 83–88, 2019. [20] S. N. H. S. Abdullah, K. Omar, A. S. Zaini, M. Petrou, and M. Khalid,
[7] J. Tang, L. Wan, J. Schooling, P. Zhao, J. Chen, and S. Wei, “Automatic “Determining adaptive thresholds for image segmentation for a license
number plate recognition (ANPR) in smart cities: A systematic review plate recognition system,” International Journal of Advanced Computer
on technological advancements and application cases,” Cities, vol. 129, Science and Applications, vol. 7, no. 6, 2016.
p. 103833, 2022. [21] D. Harjani, M. Jethwani, N. Keswaney, and S. Jacob, “Automated
[8] S. Cao, Y.-J. Wu, and F. Jin, “New Radar Sensor Technology for parking management system using license plate recognition,”
Intelligent Multimodal Traffic Monitoring at Intersections,” 2021. International Journal of Computer Technology and Applications, vol. 4,
[9] K. T. Islam et al., “A vision-based machine learning method for barrier no. 5, p. 741, 2013.
access control using vehicle license plate authentication,” Sensors, vol. [22] W. Wang, J. Yang, M. Chen, and P. Wang, “A light CNN for end-to-end
20, no. 12, p. 3578, 2020. car license plates detection and recognition,” IEEE Access, vol. 7, pp.
[10] R. Balia, S. Barra, S. Carta, G. Fenu, A. S. Podda, and N. Sansoni, “A 173875–173883, 2019.
deep learning solution for integrated traffic control through automatic [23] I. V. Pustokhina et al., “Automatic vehicle license plate recognition
license plate recognition,” in Computational Science and Its using optimal K-means with convolutional neural network for intelligent
Applications–ICCSA 2021: 21st International Conference, Cagliari, transportation systems,” Ieee Access, vol. 8, pp. 92907–92917, 2020.
Italy, September 13–16, 2021, Proceedings, Part III 21, Springer, 2021, [24] M.-X. He and P. Hao, “Robust automatic recognition of Chinese license
pp. 211–226. plates in natural scenes,” Ieee Access, vol. 8, pp. 173804–173814, 2020.
[11] H. O. Al-Sakran, “Intelligent traffic information system based on [25] P. Kaur, Y. Kumar, S. Ahmed, A. Alhumam, R. Singla, and M. F. Ijaz,
integration of Internet of Things and Agent technology,” Int J Adv “Automatic License Plate Recognition System for Vehicles Using a
Comput Sci Appl, vol. 6, no. 2, pp. 37–43, 2015. CNN.,” Computers, Materials & Continua, vol. 71, no. 1, 2022.
[12] Lubna, N. Mufti, and S. A. A. Shah, “Automatic number plate [26] M. M. Abdellatif, N. H. Elshabasy, A. E. Elashmawy, and M.
Recognition: A detailed survey of relevant algorithms,” Sensors, vol. 21, AbdelRaheem, “A low cost IoT-based Arabic license plate recognition
no. 9, p. 3028, 2021. model for smart parking systems,” Ain Shams Engineering Journal, vol.
[13] S.-O. Shim, R. Imtiaz, A. Siddiq, and I. R. Khan, “License Plates 14, no. 6, p. 102178, 2023.
Detection and Recognition with Multi-Exposure Images,” International [27] H. Shi and D. Zhao, “License Plate Recognition System Based on
Journal of Advanced Computer Science and Applications, vol. 13, no. 4, Improved YOLOv5 and GRU,” IEEE Access, vol. 11, pp. 10429–10439,
2022. 2023.
[14] F. Zantalis, G. Koulouras, S. Karabetsos, and D. Kandris, “A review of [28] G. Jocher et al., “ultralytics/yolov5: v5. 0-YOLOv5-P6 1280 models,
machine learning and IoT in smart transportation,” Future Internet, vol. AWS, Supervise. ly and YouTube integrations,” Zenodo, 2021.
11, no. 4, p. 94, 2019.
[29] S. Raj, Y. Gupta, and R. Malhotra, “License plate recognition system
[15] H. Song, H. Liang, H. Li, Z. Dai, and X. Yun, “Vision-based vehicle using yolov5 and cnn,” in 2022 8th International Conference on
detection and counting system using deep learning in highway scenes,” Advanced Computing and Communication Systems (ICACCS), IEEE,
European Transport Research Review, vol. 11, no. 1, pp. 1–16, 2019. 2022, pp. 372–377.
[16] B. S. Abunasser, M. R. J. AL-Hiealy, A. M. Barhoom, A. R. Almasri, [30] Bangyal, W.H., Nisar, K., Soomro, T.R., Ag Ibrahim, A.A., Mallah,
and S. S. Abu-Naser, “Prediction of instructor performance using G.A., Hassan, N.U. and Rehman, N.U., 2022. “An Improved Particle
machine and deep learning techniques,” International Journal of Swarm Optimization Algorithm for Data Classification”. Applied
Advanced Computer Science and Applications, vol. 13, no. 7, 2022. Sciences, 13(1), p.283.

1018 | P a g e
www.ijacsa.thesai.org