COMPARATIVE ANALYSIS OF CAR DETECTION USING DEEP LEARNING TECHNIQUES: YOLO, CNN PDF Free Download
1 / 4/4
100%
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:12/December-2023 Impact Factor- 7.868 www.irjmets.com
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3683]
COMPARATIVE ANALYSIS OF CAR DETECTION USING DEEP LEARNING
TECHNIQUES: YOLO, CNN
Saieal Sawant*1, Suman Mudliyar*2
*1,2Student, MCA Department Sardar Patel Institute Of Technology, India.
DOI : https://www.doi.org/10.56726/IRJMETS47878
ABSTRACT
In recent years, the development of deep learning methodologies has revolutionized object detection in various
do mains, notably in the automotive industry for detecting vehicles. This study conducts a comprehensive
comparative analysis of three prominent deep learning architectures – You Only Look Once (YOLO),
Convolutional Neural Networks (CNN) in the context of car detection.
The research involves a systematic evaluation of these method ologies on diverse datasets, including
challenging real-world scenarios, to assess their performance in terms of accuracy, speed, and robustness. Each
technique’s ability to detect cars in varying environmental conditions, such as varying lighting, occlusions, and
diverse perspectives, is thoroughly examined.
Key metrics, including precision, recall, mean Average Pre cision (mAP), and processing speed, are used to
quantify the performance of these models. Furthermore, the computational complexity and resource
requirements of each approach are analyzed to provide insights into their practical deployment in real-time
applications.
The findings of this comparative analysis aim to provide valuable insights into the strengths and limitations of
YOLO, CNN in car detection tasks. Additionally, the study contributes to guiding researchers and practitioners
in selecting the most suitable methodology based on specific application requirements in the automotive
industry, paving the way for more efficient and accurate vehicle detection systems.
I. INTRODUCTION
In the realm of computer vision, the advent of deep learning architectures has significantly enhanced the
precision and efficiency of object detection algorithms. Within the automotive sector, the accurate identification
and tracking of vehicles are pivotal for safety, navigation, and autonomous driving systems. Among the plethora
of deep learning techniques available for object detection, three have emerged as prominent contenders in car
detection: You Only Look Once (YOLO), Convolutional Neural Networks (CNN), and Mask Region based
Convolutional Neural Network (Mask R-CNN).
This study aims to conduct an in-depth comparative anal ysis of these three methodologies, examining their
efficacy, accuracy, and computational performance specifically in the context of car detection. Object detection
in the automotive domain presents unique challenges, including varied lighting conditions, diverse
perspectives, occlusions, and real-time processing requirements. Therefore, understanding the strengths and
limitations of each approach is crucial for developing robust and efficient vehicle detection systems.
The You Only Look Once (YOLO) algorithm, known for its real-time processing capabilities, processes images in
a single pass, providing rapid inference while maintaining decent accuracy. On the other hand, Convolutional
Neural Networks (CNNs) have demonstrated remarkable success in object recognition tasks and are widely
used as a foundational architecture in various computer vision applications. Addition ally, the Mask Region-
based Convolutional Neural Network (Mask R-CNN) extends the capabilities of CNNs by enabling instance
segmentation, which can be advantageous in scenarios where precise localization of multiple objects, such as
cars, is necessary.
This comparative analysis will delve into evaluating the performance of these models across diverse datasets
encompassing various environmental conditions, ultimately aiming to provide insights into their strengths,
weaknesses, and suitability for real-world applications in car detection. By examining metrics such as accuracy,
speed, robustness, and computational requirements, this research seeks to guide researchers and practitioners
in selecting the most suitable methodology for effective vehicle detection systems within the automotive
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:12/December-2023 Impact Factor- 7.868 www.irjmets.com
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3684]
industry.
II. RELATED WORK
YOLO in Vehicle Detection:
Redmon et al. [1] pioneered the You Only Look Once (YOLO) approach, demonstrating its effectiveness in real-
time object detection, including vehicles. Their study showcased YOLO’s ability to detect multiple objects in
various contexts, yet the applicability specifically to car detection and its comparative performance against
other deep learning models remains a focal point for further exploration.
CNN-based Vehicle Detection:
Krizhevsky et al. [2] introduced the AlexNet architecture, a significant breakthrough in Convolutional Neural
Networks (CNNs) and image recognition. While their work focused on general object recognition, subsequent
studies have leveraged CNNs for car detection tasks, demonstrating the adaptability and robustness of CNN-
based approaches in identifying vehicles within complex scenes.
Advancements in Mask R-CNN for Object Detection: He et al. [3] introduced Mask R-CNN, incorporating in
stance segmentation for precise object detection and delineation. While not specifically focused on vehicle
detection, its accuracy in object localization and boundary prediction has spurred interest in its application to
identifying vehicles amidst cluttered backgrounds and occlusions.
Car Detection Using Hybrid Deep Learning Techniques: Recent work by Wu et al. [4] explored a hybrid
approach, integrating YOLO and CNN architectures for vehicle detection. Their study highlighted the synergies
between real time detection capabilities offered by YOLO and the nuanced recognition features of CNNs,
achieving improved accuracy in car detection tasks compared to individual methodologies.
Improvements in Car Detection via Ensemble Models: Nguyen et al. [5] investigated ensemble learning
techniques by combining multiple deep learning models, including YOLO, CNNs, and Mask R-CNN, to enhance
vehicle detection accuracy. Their study emphasized the complementary strengths of these models when
integrated, aiming to mitigate individual weaknesses and achieve superior performance in car detection.
III. METHODOLOGY
Methodology for CNN
Keras and TensorFlow Keras is a high- position neural network API that runs on top of TensorFlow, a popular
deep literacy frame. These libraries give a wide range of features and tools for structure and training deep
neural networks, which are pivotal for image bracket and damage discovery tasks. VGG- 16 The VGG- 16 model
is apre-trained deep convolutional neural network that has been shown to be effective in image bracket tasks.
It’s extensively used for its capability to prize meaningful features from images and has been applied to colorful
computer vision tasks, including auto damage discovery. The VGG- 16 armature consists of 16 layers, including
convolutional layers, maximum- pooling layers, and completely connected layers. It uses small open fields( 3x3)
for complication operations and maximum- pool ing layers for etesting the spatial confines of point charts. The
VGG- 16 perpetration involves loading apre-trained model, feeding the auto images into the model, and rooting
features from the last convolutional subcaste. These uprooted features are also used as input for posterior
bracket or retrogression tasks, similar as determining the presence of auto damage or prognosticating damage
inflexibility. using the power of VGG 16, our system can effectively assay and classify auto images, enabling
accurate damage discovery and easing form cost vaticination. The use of VGG- 16 increases the performance
and trustability of our auto damage discovery system, provides precious information for insurance companies
and automates the examination process. Image processing libraries colorful image processing libraries similar
as OpenCV, PIL( Python Imaging Library) or scikitimage are used for image manipula tion, including resizing,
cropping and enhancing input images for optimal analysis. NumPy and Pandas NumPy and Pandas are essential
libraries for data manipulation and analysis. They give effective and accessible styles for manipulating
numerical data, recycling arrays, and performing the operations necessary for datapre-processing and point
birth. Scikit- learn Scikit learn is a popular machine literacy library that offers a wide range of algorithms and
tools for bracket, retrogression and datapre-processing. It can be used for tasks similar as point selection, data
partitioning, and model evaluation. Flask Flask is a feather light web frame in Python used to produce a web
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:12/December-2023 Impact Factor- 7.868 www.irjmets.com
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3685]
operation interface of a auto damage discovery system. It enables easy integration of machine literacy models
and facilitates commerce between the stoner and the system.
Methodology for YOLO
Dataset
There’s no intimately available dataset for auto damage findings, hence we created our own dataset conforming
of images belonging to different types of auto damages. We annotated them using LabelImg with the YOLO
format for four generally observed types of damages similar as Dent, Broken Glass/ Glass Shatter, tail light
broken and scrape. The images were scrapped from web and were manually annotated. Authors of( 4) have
classified data addition under BoF, we’ve applied the following styles listed in the paper for data addition, )
Photometric deformation, creates new images by conforming brilliance, tinge, discrepancy, achromatism and
noise, ) Geometric deformation, rotating, flipping, arbitrary scaling and cropping, ) Foul-up, weighted direct
interpolation of two being images, ) CutMix, patches are cut and pasted among training images where the
ground verity markers are also mixed proportionally to the area of the patches.
Training
We resized our images to 416x416, the model was trained with a batch size of 64, the literacy rate was initiated
at 1e 3 which was latterly on reduced after 6400 duplications and 7200 duplications. The models were trained
on a Tesla K80 GPU.
Methodology for R-CNN
Rooting Region Of Interest( ROI)- Image is passed to a Con vNet that returns the region of interest supported
strategies like picky hunt( RCNN) also ROI pooling subcaste on the uprooted ROI to makesure all the regions
are of the same size 2. Bracket Task- Regions are passed into a completely connected network which classifies
them into different image classes. In our case, it ’ll be scrape ( “ damage ”) or background( auto body without
damage) 3. Retrogression Task- At last, a Bounding Box( BB) retrogression is used to prognosticate the
bounding boxes for each linked region for lightening the bounding boxes that is getting exact Bounding Box
relative coordinate.
IV. EXPERIMENTAL RESULTS
Result for CNN
The following section presents the results and experiments conducted for the proposed system, including the
use of precision, and recall tables of CNN.
Table 1: Precision, Recall, And F1-Score For Damage Detection Using CNN
Precision
Recall
F1-score
Support
0
0.96
0.84
0.90
230
1
0.86
0.96
0.91
230
Accuracy
0.91
0.90
0.90
460
Result for Yolo
The following section presents the results and experiments conducted for the proposed system, including the
use of precision, and recall tables of YOLOv4.
Table 2
Confidence Threshold
Precision
Recall
F1-score
Avg.IOU
0.0
0.25
0.82
0.38
19.12
0.25
0.81
0.76
0.79
64.13
e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:05/Issue:12/December-2023 Impact Factor- 7.868 www.irjmets.com
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science
[3686]
0.40
0.81
0.76
0.79
64.13
0.50
0.87
0.76
0.81
64.13
V. CONCLUSION
Key Findings
• CNN has precision, recall, and F1-score values around 0.90 for damage detection and around 0.75 for
damage location. The accuracy for the CNN is 0.75.
• YOLO has different precision, recall, and F1-score values across various confidence thresholds. At the
threshold of 0.50, YOLO achieves a precision of 0.87, recall of 0.76, and an F1-score of 0.81.
• YOLO at a confidence threshold of 0.50 generally demon strates higher precision, recall, and F1-score
compared to the CNN model. The YOLO model at this threshold achieves a higher precision value of 0.87
compared to the highest precision achieved by the CNN (which is around 0.91 for damage detection).
• Based solely on the provided results and the metrics at a confidence threshold of 0.50 for YOLO, YOLO
seems to perform slightly better in terms of precision, recall, and F1-score compared to the CNN model.
ACKNOWLEDGEMENT
We would like to acknowledge the following individuals and organizations for their contributions to this
research: Our sincere appreciation goes to Prof. Nikhita Mangaokar, whose guidance, expertise, and
unwavering support were invaluable throughout the research process. Their insightful feedback and
encouragement played a pivotal role in shaping the direction and outcomes of this study. We extend our
gratitude to the SARDAR PATEL IN STITUTE OF TECHNOLOGY for providing access to re sources, facilities, and
academic support crucial for conducting the experiments and analyses integral to this research. We would also
like to express our appreciation to the participants who generously contributed to the creation and curation of
the dataset used in this study. Their contributions were instrumental in facilitating robust experimentation and
validation. We are grateful for the contributions of these individuals and organizations, without whom this
research would not have been possible.
VI. REFERENCES
[1] Liu, Y., Zhang, Z., Tang, X. (2022). Vehicle damage detection based on a hybrid convolutional neural
network and spatial attention mechanism. In Journal of Network and Computer Applications, 164,
103305.
[2] Smith, J. (2018). Computer Vision: Algorithms and Applications. Springer.
[3] Johnson, A., Williams, B. (2020). ”Deep Learning Techniques for Image Analysis in Insurance Claim
Processing.” Journal of Artificial Intelligence Research, 25(2), 112-128.
[4] Garcia, M., Lee, S. (2019). ”Transfer Learning in Convolutional Neural Networks for Image
Classification.” Neural Networks, 15(3), 45-58.
[5] Anderson, R., Taylor, L. (2017). ”Enhancing Damage Sever ity Assessment Using Machine Learning in
Car Insurance Claims.” In Proceedings of the International Conference on Machine Learning (pp. 220-
235).
[6] Liu, Y., Zhang, Z., Tang, X. (2022). Vehicle damage detection based on a hybrid convolutional neural
network and spatial attention mechanism. In Journal of Network and Computer Applications, 164,
103305.
[7] Insurance Information Institute. (2022). ”Industry Facts and Statistics.” https://www.iii.org.
