Proceedings of the 2023 International Conference on Software Engineering and Machine Learning

DOI: 10.54254/2755-2721/8/20230109

Image detection and segmentation using YOLO v5 for

surveillance

Mohanapriya S1,2, Mohana Saranya S1, Kumaravel T1, Sumithra P1

1
Department of Computer Science and Engineering, Kongu Engineering College, Perundurai,
Erode, Tamil Nadu, India

2
mohanapriyas.cse@kongu.edu

Abstract. Segmentation an advancement of object detection where bounding boxes are placed
around object in object detection whereas segmentation is used to classify every pixel in the
given image. In Deep Learning, Yolov5 algorithm can be used to perform segmentation on the
given data. Using YOLOv5 algorithm objects are detected and classified by surrounding the
objects with the bounding boxes. Compared to the existing algorithms for segmentation,
YOLOv5 algorithm has improved time complexity and accuracy. In this paper YOLOv5
algorithm is compared with the existing CNN algorithm.

Keywords: deep learning, object detection, segmentation, YOLOv.

1. Introduction

1.1. Deep learning

Artificial Intelligence is the one which is used for learning from any unsupervised data that doesn't
contain predetermined labels[1]. Making computers think and behave like humans is the idea behind
Artificial Intelligence. Deep learning part of the Artificial Intelligence is used to carry out many
complicated tasks like recognition of patterns, object detection, segmentation like semantic and
instance segmentation. All the functionalities of Deep Learning are carried with huge amount of data
which ranges from Gigabytes to Terabytes. With the help of deep learning valuable information can be
retrieved from large amount of data [2].

1.2. Object detection

Object detection is crucial in computer vision, automated vehicles, and industrial automation and other
applications. It's difficult to detect items in real time. Deep learning outperforms traditional target
detection in object detection. This paper provides an object detection and segmentation system that is
conceptually simple, versatile, and broad. The method proposed in this paper successfully detects
objects in images and videos while producing high-quality segmentation for each instance also. To
detect objects, the YOLO algorithm requires only one forward propagation across a neural network.
The You Only Look Once model family is a collection of end-to-end deep learning models for quick
object detection. The newest algorithm in the YOLO family is YOLOv5. This paper discusses various
channel characteristics. It also improves the detecting approach so that more detailed feature

© 2023 The Authors. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0
(https://creativecommons.org/licenses/by/4.0/).

160
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230109

information may be preserved. The final findings suggest that the upgraded YOLO V5 approach
enhances performance.[3]

1.3. Segmentation
Segmentation an advancement of object detection is used to detect and classify objects in the image.
Instance segmentation is applied in many real time applications like self driving cars, agriculture,
medical systems etc. CNN one of the important object detection frameworks is used for detecting
objects in the image. All the object detection and segmentation frameworks were developed based on
this CNN algorithm. One such detection algorithm is the YOLOv5 algorithm. YOLOv5 algortihm is
proved to be the state-of-the-art algortihm for segmentation of objects in the image.[4]

2. Literature review
Image segmentation has gotten a lot of attention recently as one of the successful applications of
categorising objects and applying masks for the item present in the image. The study addresses over a
hundred deep learning-based segmentation algorithms proposed through 2019 and includes the most
recent research on instance segmentation [5]. We present a thorough examination and analysis of
several elements of these approaches, including training data, network architecture selection, loss
functions, training state, and major contributions. We give a comparison of the performance of the
approaches under consideration, as well as many obstacles and possible future directions for deep
learning-based instance segmentation models. YOLO established a single unified architecture for
breaking go picture into bounding boxes and calculating class probabilities for each box, in
comparison to object identification approaches that came before it, such as R-CNN. As a result,
YOLO was able to execute significantly faster and with greater precision. It may also properly
anticipate artwork. [6]
Object Detection aims to construct a general object recognition network, complex degradation
methods including noise, blurring, rotating, and cropping of images were applied. The model's
generalisation and robustness were improved by employing degraded training sets during training. The
study found that the model's generalisation and resilience when used on damaged images were weak
when trained on standard sets. After training the model with damaged images, average accuracy
increased. It was demonstrated that the wide degenerative model outperformed the conventional model
in terms of average accuracy for degraded images.
The YOLO Network Model says an improved network model is developed and a new network
structure known as YOLO-R has been proposed to boost the network's capacity to extract information
from superficial pedestrian characteristics by including pass through layers into the original YOLO
network. The INRIA data collection's test set had been used to assess YOLO v2 and YOLO-R network
models. Compared to YOLO v2 network model , YOLO-R network model performs better. The real-
time performance criterion was met when the detecting frame rate increased to 25 frames per second.
Solder Joint Recognition and Detection in Automotive Door Panels, a solder joint recognition
method based on the YOLO algorithm that gives the kind and location of solder joints in real time for
automobile door panels. In order to more easily identify tiny patch crossings, this study applies the
YOLO approach, which employs staggered forecasts, expecting on many size highlight guides, and
merging the expectation outcomes to form the final conclusion. The proposed YOLO approach
successfully locates solder connections in real time. This increases the productivity of the production
line and is crucial for the flexible and real-time welding of vehicle door panels.
Though many works have been proposed to address the problem of object detection and
segmentation, still a research gap available to improve accuracy in this area. This paper focuses to use
YOLO v5 algorithm for object detection and segmentation to improve that gap.

161
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230109

3. Proposed work

3.1. YOLO V5 algorithm

YOLO algorithm is known for its high performance and quick time requirements. It is one of the most
popular deep convolutional neural methods for object segmentation. The PyTorch framework is used
in YOLOv5. It is the most recent version of the YOLO object recognition model, which was created
with the help of 58 open source contributors throughout time. Other deep neural networks may be used
to detect things as well. One of them is the Mask-RCNN [9], which is designed to handle the problem
of instance segmentation in computer vision machine learning. Mask-RCNN is more exact, but it takes
longer to process. YOLO and Mask R-CNN models give results of high recall and precision for
detecting a ball sports. The YOLOv5 network is used in this paper since it is a good and quicker
detector with excellent levels of performance. Other architectures, such as the MaskRCNN, may be
able to achieve comparable detection results while providing more exact object positioning.

3.2. Working principle of YOLO V5

In YOLO algorithm, the image is divided into ‘n’ grids of equal size. In each grid, the object contained
in that grid is detected and localized. The grids are responsible for predicting coordinates of bounding
box according to their cell coordinates. Prediction in this way greatly reduces computation of detection
and recognition buts it leads to duplicate predictions. This issue is dealt with maximal suppression in
YOLO
YOLO eliminates the bounding boxes that have probability score very minimum. This is done by
seeing the score of each decision and finding out which one is the largest. After finding the largest
value, YOLO eliminates all the bounding boxes which have the highest IoU value with the current
bounding box which have the high value. The above step is repeated until the target bounding box is
obtained in Figure 1.

Input Images

Image Segmentation using Neural Network

Image Extraction

Yolov5 Algortihm

Detecting & locating the objects in image

Figure 1. Flow chart of proposed work.

162
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230109

3.3. Model backbone

From the raw photos provided, Model Backbone is used to extract significant characteristics. To
extract highly useful data from an input image, Cross Stage Partial Networks (CSP) can be employed
as the backbone.

3.4. Model neck

The fundamental purpose of a model neck is to build feature pyramids. Models generalise easily to
objects of various sizes thanks to feature pyramids. The ability to recognise the same thing in various
sizes and scales is helpful. Models that employ feature pyramids perform well on unobserved data.
PANet is utilised in Yolo V5 as a neck to get feature pyram.

3.5. Model head

The last detecting step is carried out by the model Head. It applies anchor boxes to features and creates
a final output vector with bounding boxes, an objectness score, and a class likelihood.

3.6. Object detection using neural network

Using a neural network classifier with a feed-forward, one hidden layer network and back propagation
as the learning method, an object detection algorithm is built. The definition of efficient object
characteristics, which are utilised to train the classifier, is a crucial component of this system.

4. Results and discussion

4.1. Dataset
A COCO dataset of nearly 10-20 lakhs that has already been trained by using predefined functions is
used to assess the proposed work. The dataset images are frame-by-frame trained. From the COCO
dataset, we took 5000 images for testing. The MS COCO dataset offers a sizable dataset for object
recognition and instance segmentation, both of which were used to test several deep learning
techniques. Figure 2 demonstrate an example input image and output from the dataset.

Figure 2. Input and output image.

4.2. Accuracy
Accuracy is used to measure how the model performs for different classes of objects. It is the ratio
between total number of correct predictions to the total number of predictions made. The Yolo V5 and
CNN algorithms' degrees of accuracy are displayed in Table 1.

163
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230109

Table 1. Accuracy level (%) of Yolo V5 and CNN.

No. of images Time Complexity of Yolo Time Complexity of CNN(ms)
V5(ms)
1000 0.22 1.36
2000 0.14 2.20
3000 0.40 2.207
4000 0.10 1.94

Figure 3 shows the accuracy level comparison of Yolo V5 and CNN. In the figure, we can see that
Yolo V5 performs better than CNN.

Figure 3. Accuracy level comparison of Yolo V5 and CNN.

4.3. Time complexity

The amount of time taken to run an algorithm is known as the Time complexity. It is known as the
Computational Complexity. It can be measured in ms. The Time complexity of the Yolo V5 and CNN
algorithms are shown in Table 2.

Table 2. Time complexity (%) of Yolo V5 and CNN.

No. of images Accuracy of Yolo V5(%) Accuracy of CNN (%)

1000 93 37
2000 54 43
3000 84 18
4000 95 25

The comparison of time complexity between Yolo V5 and CNN is shown in Figure 4. The figure
shows.

164
 Proceedings of the 2023 International Conference on Software Engineering and Machine Learning
DOI: 10.54254/2755-2721/8/20230109

Figure 4. Time complexity comparison of Yolo V5 and CNN.

5. Conclusion and future work

The segmentation technique an advancement of object detection is used to detect and classify pixels in
the image. The YOLOv5 method, which is based on deep learning and excels at object detection, has
been made available. Yolov5 significantly reduces time complexity and improves segmentation
accuracy when compared to earlier state-of-the-art algorithms. As a result, YOLOv5 is a superior
option for identifying things and determining objects in the image.

References
[1] Sathishkumar, V. E., Cho, J., Subramanian, M., & Naren, O. S. (2023). Forest fire and smoke
detection using deep learning-based learning without forgetting. Fire Ecology, 19(1), 1-17.
[2] Subramanian, M., Cho, J., Sathishkumar, V. E., & Naren, O. R. (2023). Multiple types of
Cancer classification using CT/MRI images based on Learning without Forgetting powered
Deep Learning Models. IEEE Access.
[3] Kogilavani, S. V., Sathishkumar, V. E., & Subramanian, M. (2022, May). AI Powered COVID-
19 Detection System using Non-Contact Sensing Technology and Deep Learning
Techniques. In 2022 18th International Conference on Distributed Computing in Sensor
Systems (DCOSS) (pp. 400-403). IEEE.
[4] Shanmugavadivel, K., Sathishkumar, V. E., Kumar, M. S., Maheshwari, V., Prabhu, J., &
Allayear, S. M. (2022). Investigation of Applying Machine Learning and Hyperparameter
Tuned Deep Learning Approaches for Arrhythmia Detection in ECG Images. Computational
& Mathematical Methods in Medicine.
[5] Krishnamoorthy, N., Prasad, L. N., Kumar, C. P., Subedi, B., Abraha, H. B., & Sathishkumar,
V. E. (2021). Rice leaf diseases prediction using deep neural networks with transfer learning.
Environmental Research, 198, 111275.
[6] Easwaramoorthy, S., Sophia, F., & Prathik, A. (2016, February). Biometric Authentication
using finger nails. In 2016 international conference on emerging trends in engineering,
technology and science (ICETETS) (pp. 1-6). IEEE.

165