Journal of Artificial Intelligence, Machine Learning and Neural Network

ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Convolutional Neural Networks for Object Detection

and Recognition

Ms. Archana Karne1, Mr. RadhaKrishna Karne2*, Mr. V. Karthik Kumar3,

Dr. A. Arunkumar4
1
UG Student, CMR Technical Campus, Hyderabad, India
2*
Assistant Professor in ECE, CMR Institute of Technology, Hyderabad, India
3
Assistant Professor in ECE, BITS Narsampet, Hyderabad, India
4
Professor in CSE, MLRITM, Hyderabad, India

Corresponding Email: 2*krk.wgl@gmail.com

Received: 15 October 2022 Accepted: 02 January 2023 Published: 04 February 2023

Abstract: One of the essential technologies in the fields of target extraction, pattern
recognition, and motion measurement is moving object detection. Finding moving objects
or a number of moving objects across a series of frames is called object tracking. Basically,
object tracking is a difficult task. Unexpected changes in the surroundings, an item's
mobility, noise, etc., might make it difficult to follow an object. Different tracking methods
have been developed to solve these issues. This paper discusses a number of object tracking
and detection approaches. The major methods for identifying objects in images will be
discussed in this paper. Recent years have seen impressive advancements in fields like
pattern recognition and machine learning, both of which use convolutional neural
networks (CNNs). It is mostly caused by graphics processing units'(GPUs) enhanced
parallel processing capacity. This article describes many kinds of object classification,
object racking, and object detection techniques. Our results showed that the suggested
algorithm can detect moving objects reliably and efficiently in a variety of situations.

Keywords: Convolutional Neural Network (CNN), Deep Learning (DL), Object Detection,
Deep Neural Networks (DNN), Object Recognition, YOLO, Object Tracking, Object
Classification.

1. INTRODUCTION

The difficult yet important task of detecting and following moving objects in a video
surveillance system. In today's society, a quick video surveillance system is necessary. The
purpose of the video surveillance system is to be focused on the identification, categorization,
and tracking of moving objects. Almost every industry, including the military, workplaces,
banks, and schools, uses moving object detection and tracking for security purposes. The use
of automatic video surveillance is crucial in the security industry[1]. the use of radar and

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 1
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

image processing technologies for the purpose of accurately identifying and following
moving objects In this study, moving object recognition and tracking are accomplished using
image processing technologies. For every object detection and tracking system, detection,
categorization, and tracking are the three crucial processes. The items of interest are initially
removed from the series of frames, separated from the backdrop, then tracked while retaining
their identity in subsequent frames. The object recognition and tracking method is applicable
to more than only video surveillance systems [2-3]; it may also be used in multimedia
databases, virtual reality, video compression, human-machine interfaces, and other areas.

The finest fundamental quality of a person is frequently their capacity for seeing. Our
capacity to see with our eyes is thought of as a gift and is crucial to how we go about our
daily lives. The fact that many visually impaired persons are constrained by their eyesight
and unable to be fully independent is a significant difficulty. People who are visually
impaired have difficulty with these activities, thus object recognition will be a crucial
characteristic they may rely on frequently. They frequently encounter difficulties moving
around and recognizing objects, especially while going on streets. The majority of visually-
capable people reach middle age at 50 [4]. A small number of applications to help the blind
have entered the market. However, there is still a lack of modernization in the "visually
impaired" people's lack of real-time continuous article acknowledgement and object
identification with speech output. With the usage of IoT[5-10], some of these apps are
focused on detecting obstacles close to the user and warning them via alarms or blaring
noises. Numerous gadgets are needed by consumers to hold for various reasons. For example,
navigation aid requires smart sticks with obstacle detectors, cell phones, navigators, etc.
These gadgets are pricey and may occasionally cause the user difficulty.

Computer vision has several exciting challenges, such classifying images and identifying
objects, both of which fall under the umbrella of object recognition. In recent years, there has
been significant scientific progress for these kinds of problems, mostly because CNN, DL
approaches, and the rise in parallelism processing power provided by GUPs have all
advanced. The goal of the image classification issue is to select a label from a predetermined
list of categories to apply to an input image. The labelling of images of skin cancer [11] and
the use of high-resolution images to detect natural disasters like floods, volcanoes, and severe
droughts while noting the impacts and damage caused by these events are just two examples
of the many practical applications and uses of this classification problem, which is central to
computer vision. The characteristics that are fed into image classification algorithms have a
critical role in how well they function [12]. This indicates that the development of machine
learning-based image categorization approaches strongly depended on the engineering of
identifying the crucial aspects of the photos that made up the database. As a result, getting
these resources has grown to be a difficult undertaking, increasing complexity and expense.
When seen as a component of the supervised learning strategy, the support vector
machines(SVM) algorithm is frequently employed for tasks including classification,
regression, and outlier identification [13]. The system's learning process for numerous objects
can be mathematically examined more easily than standard neural network design, which
makes it possible to make sophisticated changes to the algorithm with predictable results
[14]. A nonlinear separation barrier in the input space is produced by an SVM's basic
mapping of the training data to higher-dimensional feature space and separation hyperplane

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 2
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

with maximum margin [15]. Deep learning architectures with several specialized layers for
automating the filtering and feature extraction process are used by today's most reliable
object categorization and detection systems. A prediction is made, a correction is received,
and the prediction mechanism is adjusted based on the correction. At a high level, this is very
similar to how a human learns. Machine learning algorithms like linear regression, support
vector machines, and decision trees all have their own peculiarities in the learning process.
The advent of deep learning has introduced a fresh perspective on the issue, one that aimed to
address past limitations by learning abstraction in data via a stratified description paradigm
built on a nonlinear transformation [16]. Due to the widespread usage of CNN, DL-based
algorithms may learn the feature extraction stage (ConvNet or CNN). Convolution is a
specific kind of linear operation that may be viewed in this sense as the straightforward
application of a filter to a predetermined input [17]. By adjusting the convolution's
parameters, the same filter is applied to an input repeatedly to produce a feature map that
shows the positions and intensities of any discovered features. As a result, the network may
learn the ideal parameters to extract pertinent data from the database by adjusting itself to
decrease error. There have been several deep neural network (DNN)-based object detectors
developed in recent years [18-19]. In order to demonstrate how state-of-the-art DNN models
of SSD and Faster RCNN function in scientific research, the YOLO network was trained to
solve the mice tracking issue. The algorithms were taught to recognize a variety of animals in
images for the traditional detection challenge.

Object Detecting, Classification and Tracking Methods

Object Detection Methods
Finding and recognizing objects in an image or video sequence is the challenge of object
detection. Even when partially obscured from vision, objects may still be identified.
Computer vision systems are still having trouble with this task. For object detection, several
techniques have been developed throughout the years.

Background Subtraction Method

The foreground of an image is retrieved for further processing using the technique known as
background subtraction, which is also known as foreground detection. When morphological
object localization is needed as post processing following the step of picture preprocessing,
which includes image noise removal, this approach may be used. A common technique for
identifying moving objects in films taken with a stationary camera is background removal.
Backdrop images or background models are used to identify moving objects by comparing
the differences between the current frame and a reference frame. The approach of background
subtraction is weak at blocking interference and sensitive to changes in the environment.

Optical Flow Method

The distribution of the objects' apparent velocities inside a picture is known as optical flow.
The velocities of the objects in the movie may be calculated by calculating optical flow
between video frames. In general, moving objects will appear to move more visibly the closer
they are to the camera and the faster they are travelling than farther away. For motion-based
object recognition and tracking systems, optical flow estimation is used in computer vision to
characterize and quantify motion of objects in video streams.

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 3
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Frame Differencing Method

By using a technique called frame differencing, the computer can determine if two video
frames differ from one another.

Single Shot Multibox Detector (SSD)

One of the best detectors in terms of speed and accuracy is the SSD[37], which uses
convolutional filter applications and feature map extraction as its two primary processes for
object detection. The VGG-16 network [38], on which the SSD design is based, has great
performance in high-quality image classification tasks and enjoys widespread use in
challenges involving transfer learning. The convolutional kernel application to an input
picture in the SSD architecture is shown in Figure 1. A collection of auxiliary convolutional
layers replace the initial VGG fully connected layers in the model, enabling the extraction of
features at various scales and progressively reducing the size of the input to each succeeding
layer. The use of precomputed, fixed-size bounding boxes known as priors to the initial
distribution of ground truth boxes is taken into account during the bounding box creation.
These priors are chosen to maintain an intersection over union (IoU) ratio of 0:5 or above.

Figure1. To the end of the basic network, the SSD network contains various feature levels.

Faster region convolutional neural network (Fast RCNN)

Another cutting-edge CNN-based deep learning object identification method is the Fast
RCNN [20]. In this design, a convolutional network is used to create a convolutional feature
map from the input picture. A different network is utilized to learn and forecast these regions
rather than utilizing the selective search algorithm to detect the region recommendations
made in earlier rounds [21-22]. A region of interest (ROI) pooling layer is then used to
reshape the projected region proposals, categorize the image within the proposed region, and
forecast the offset values for the bounding boxes. The region proposal network (RPN)
training technique uses a binary label for each anchor, with one denoting an object's existence
and zero denoting its absence. According to this strategy, any IoU over 0:7 identifies an
object's presence, while any value below 0:3 denotes an object's absence. The unified
network for object detection used in the Faster RCNN architecture is shown in Figure 2. The
region proposal network module instructs the Fast RCNN module where to seek using the
currently fashionable language of neural networks with "attention" processes [23].

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 4
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Figure2. For object detection, faster RCNN functions as a single, integrated network [20].

Object Classification Method

The object is classed by assigning each object to a class based on characteristics once the
object of interest has been identified from a sequence of frames. Features are defined in a
broad variety of ways. The feature specifies the target image's form, size, colour, and motion
for this purpose. The following features are employed in object tracking.

Edges- Image intensities frequently undergo significant fluctuations at object borders. This
change's identification is done using edge detections. The fact that edges are less responsive
to changes in light than colour characteristics is a significant characteristic of edges [24].

Motion- A major indication for classifying moving objects has been the periodic behaviour
of non-rigid articulated item motion [25]. Object motion may also be tracked using optical
flow.

Color- All video frame formats are built on a concept of several colour spaces. Different
frame data can be recorded in several colour formats, including grayscale, RGB, YCbCr, and
HSB. The letters read (R), green (G), blue (B), or RGB are used to indicate colour pictures.

Texture- Texture is utilized to help identify the subject or object of interest. It evaluates
characteristics like smoothness and regularity of a surface by measuring intensity variation of
that surface.

Object Tracking Method

Target tracking looks for an object's location in each each frame of video in order to build the
root for that object above time. There are three types of object tracking: silhouette-based
tracking, point tracking, and kernel tracking.

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 5
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Point Tracking Method - Veenman created point tracking, a dependable, strong, and precise
tracking technique. Their feature points serve as a representation for moving objects. Point
tracking techniques are further separated into two groups: deterministic and statistical. The
foundation of object tracking is a point that is depicted in an object that is detected in a
subsequent frame, and the association of points is based on the prior state of the item. To
identify an item in every frame, an external mechanism is needed.

Kernel Tracking - Kernel describes an object's representation of its elliptical or rectangular

form and appearance. From frame to frame, the object's motion is calculated. In succeeding
frames, the object's motion is represented by parametric motion or dence flow field
computation. Simple template matching, mean shift technique support vector machines, and
layering-based tracking are further categories for kernel tracking. Tracking may be done for a
single item in a video using basic template matching, which verifies a reference picture with a
frame that has been taken out of the film. Tracking the motion of the item is done using
translation and scaling, and the object of interest is defined using a rectangular frame. After
that, the tracked item and backdrop are separated. SVMs need a training set of values.
Positive values are used to contain these, whereas negative values are used to contain targets
that are not being monitored. In layering-based tracking, many objects may track.

Silhouette Tracking - Tracking using silhouettes is utilized when a whole object region is
needed. Numerous items, such as the human body, head, and hand, have complicated
geometries that may be correctly represented using a silhouette-based technique. By utilizing
an object model created from previous frames, the silhouette tracker seeks out the object
region in each frame. Shape matching and Contour tracking are two subcategories of
silhouette tracking.

You Only Look Once (YOLO) Algorithm

YOLOis a cutting-edge object identification method designed for real-time applications; in
contrast to some of its rivals, it is not a conventional classifier used for object detection[26].
In order for YOLO to function, the input picture is split up into a grid of SxS cells, where
each cell is in charge of five bounding box predictions that characterize the rectangle around
the item. Additionally, it generates a confidence score, which expresses how confidently an
object was contained. Therefore, just the form of the box affects the score; the type of object
in the box has no bearing on it. As with a typical classifier, a class is predicted for each
anticipated bounding box, yielding a probability distribution over all potential classes[34-36].
One final score that indicates the likelihood that each box contains a certain type of item is
created by combining the bounding box confidence score with the class prediction score. Due
to these design decisions, the majority of the boxes will have low confidence ratings; thus,
only the boxes with final scores that are higher than a certain threshold are maintained. How
the YOLO network processes a picture is shown in Figure 3. The input is first processed
through a CNN, which creates the bounding boxes with its viewpoints' confidence ratings and
creates the class probability map. The final forecasts are created by combining the outcomes
of the earlier processes.

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 6
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Figure3. Model detection for YOLO as a regression problem[27].

YOLO is an engaging article openness evaluation. Even if it isn't currently the most cautious
article straightforwardness figure, it is a fantastic option when an unsurprising, undeniable
need is needed without losing a significant amount of accuracy. YOLO employs a lone CNN
to organize things utilizing swaying boxes for both gathering and limiting them. YOLOv2
provides high accuracy and dependable getting ready, but it has more confinement errors and
a lower survey response than other area-based pointer checks. A resurrected version of
YOLO, known as YOLOv2, beats the lower study response and produces accuracy with
vivacious openness. The improvements in YOLOv2 are quickly examined under: The
completely related layers that were submitted for the cutoff box expectation were discarded.
To change the spatial yield of the framework from 7x7 to 13x13, one pooling layer has to be
removed. Since classifiers anticipated that yield names would be fully distinct, the yield
object classes were mostly unimportant. YOLO has the capacity to convert the
aforementioned scores into probabilities as much as possible. YOLOv3 has a multi-name
strategy. A score that is extraordinary can be seen in non-prohibitive yield inscriptions.
Instead of utilizing the soft max work, YOLOv3 enrolls the likelihood of the data having a
location with a certain cutting by applying free essential classifiers. YOLOv3 determines the
framework disappointment using a scene that is shaped using cross-entropy for each name
rather than the mean settled slip. Maintaining a few essential charming strategies from the
soft max work reduces the complexity of the check. A common, quicker, and more grounded
version of YOLO is YOLO-9000. The YOLO algorithm was initially proposed by Joseph
Redmon and his colleagues. In 2015, he released a paper on YOLO with the working title
"You Only Look Once" Real-Time item recognition, and it became an instant hit. CNN is
followed by YOLO. When making predictions, the algorithm only "looks once" at the image
since there is only one propagation that occurs throughout the neural network. Compared to
other methods of object identification, the YOLO model is the fastest and most effective. The
main benefit of YOLO is its quickness. There are 45 frames per second in this. The model is
constructed in a concise manner that allows its network to become used to abstract
descriptions of things [28].

Evaluation Parameters
In this paper, we evaluated the effectiveness of the moving object identification method using
the precision (P), recall (R), and F1 measures. The performance of the object identification
model was assessed using the mean average precision (mAP). Based on the true category and

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 7
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

the detection category, the detection results were split into four cases: true positive (TP), false
positive (FP), false negative (FN), and true negative (TN).
TP
Precision =
TP  FP
TP
Recall =
TP  FN
2 PR
F1=
PR

2. RESULTS

The use of the SSD and RCNN object detection techniques is demonstrated using a portion of
the PASCAL VOC [29] dataset. Six classes, out of the 20 offered, were chosen as a sample.
The sample size used for each class is shown in Figure 4. The dataset's photos were randomly
divided into 1911 for training, which corresponds to 50%, 1126 for validation, which
corresponds to 25%, and 1126 for test, which also corresponds to 25%. The dataset utilized
for the YOLO network published in [30] was also examined to further highlight the uses of
these algorithms in academic study. The dataset, which is presented in [30], is made up of
photographs from three studies that involve behavioural trials on mice. The sample size
chosen from each of the datasets utilized in this work is shown in Figure 6. A total of 3707
frames from a top view of the mouse social interaction experiment chamber were utilized for
the ethological evaluation [31]. A sample of 3073 frames was chosen for the automated
home-cage [32] from a side perspective of behavioural studies. A selection of 6842 frames
from the CRIM13 [33], including 3492 from the side view and 3350 from the top view, were
chosen.

Figure 4. Description of the SSD and RCNN network datasets

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 8
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Figure5. Examples of the SSD (a-d) and Faster RCNN(e-i) networks' output

Figure 6. An explanation of the dataset used with the YOLO network

More results on the effectiveness of object detection is shown in Figure 7. In the beginning, it
displays the mean average precision, which is the mean value of the average precisions for
each class. Average precision is the average value of 11 points on the precision-recall curve
for each potential threshold, or all the probabilities of detection for the same class. We
examined the models' performance in terms of accuracy, speed, and model size, shown in
figure 8.

Figure 7. Results of the mean average Precision after 100 training iterations

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 9
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

Figure 8. Comparisons of different methods

3. CONCLUSION

In this paper, we presented a novel approach to the detection and identification of moving
objects, and also an overview of the related literature as well as an object tracking literature
study are presented in this paper. The three kinds of tracking techniques are object detection,
object classification, and object tracking.The future of object detection offers tremendous
prospects in a variety of businesses. Based on the resources available, methods for object
detection and video processing are presented.The experiment results of CNN, SVM, SSD,
FPN, YOLO2 and YOLO3, were compared by means of precision, recall, F1-score, average
precision and generating frames per sec.Results of the experiments show that the suggested
approach accurately and successfully detects moving objects.As future work,, we will
concentrate on developing a moving object detection algorithm that is more reliable and
integrating it into embedded surveillance application systems.

4. REFERENCES

1. Kumar, A. Arun, and Radha Krishna Karne. "IIoT-IDS Network using Inception CNN
Model." Journal of Trends in Computer Science and Smart Technology 4.3 (2022): 126-
138.
2. Karne, RadhaKrishna, and T. K. Sreeja. "ROUTING PROTOCOLS IN VEHICULAR
ADHOC NETWORKS (VANETs)." International Journal of Early Childhood 14.03:
2022.
3. Vaigandla, Karthik Kumar, Sravani Thatipamula, and Radha Krishna Karne.
"Investigation on Unmanned Aerial Vehicle (UAV): An Overview." IRO Journal on
Sustainable Wireless Systems 4.3 (2022): 130-148.
4. RadhaKrishna Karne, Dr Sreeja TK, “COINV-Chances and Obstacles Interpretation to
Carry new approaches in the VANET Communications” Design Engineering, 2021,
10346-10361

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 10
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

5. K. K. Vaigandla, "Communication Technologies and Challenges on 6G Networks for

the Internet: Internet of Things (IoT) Based Analysis," 2022 2nd International
Conference on Innovative Practices in Technology and Management (ICIPTM), 2022,
pp. 27-31, doi: 10.1109/ICIPTM54933.2022.9753990.
6. RadhaKrishna Karne, Dr TK. "Review On Vanet Architecture And
Applications." Turkish Journal of Computer and Mathematics Education
(TURCOMAT) 12.4 (2021): 1745-1749.
7. RadhaKrishna Karne, Dr Sreeja TK, “A Novel Approach for Dynamic Stable Clustering in
VANET Using Deep Learning (LSTM) Model” International Journal of Electrical and
Electronics Research (IJEER) , Volume 10, Issue 4, 2022, Page(s) : 1092-1098
DOI: https://doi.org/10.37391/IJEER.100454
8. Sandeep Singh Sengar, and Susanta Mukhopadhyay. “Motion Detection using Block
based Bi-directional Optical Flow Method", Journal of Visual Communication and
Image Representation, Elsevier, Vol.-49, pp. 89-103, August 2017.
9. Sandeep Singh Sengar, and Susanta Mukhopadhyay. “Moving Object Detection based
on Frame Difference and W4", Signal, Image and Video Processing, Springer, Vol.-11,
Issue-7, pp. 1357-1364, April 2017.
10. Karne, RadhaKrishna, et al. "Simulation of ACO for Shortest Path Finding Using NS2."
(2021): 12866-12873. [11]
11. Esteva A et al. Dermatologist-levelclassification of skin cancer with deepneural
networks. Nature. 2017;542(7639):115
12. Srinivas S, Sarvadevabhatla RK,Mopuri RK, Prabhu N, Kruthiventi SSS,Venkatesh
Babu R. An introduction todeep convolutional neural nets forcomputer vision. In: Deep
Learning for Medical Image Analysis. AcademicPress; 2017. pp. 25-52
13. de Menezes RST, de Azevedo Lima L,Santana O, Henriques-Alves AM, SantaCruz
RM, Maia H. Classification of micehead orientation using support vectormachine and
histogram of orientedgradients features. In: 2018International Joint Conference
onNeural Networks (IJCNN). IEEE; 2018.pp. 1-6
14. OskoeiMA, Gan JQ, Hu H. Adaptiveschemes applied to online SVM forBCI data
classification. In: 2009Annual International Conference ofthe IEEE Engineering in
Medicine and Biology Society. IEEE; 2009.pp. 2600-2603
15. Hearst MA, Dumais ST, Osuna E,Platt J, Scholkopf B. Support vectormachines. IEEE
Intelligent Systems andtheir Applications. 1998;13(4):1828
16. Pan WD, Dong Y, Wu D. Classification of malaria-infected cellsusing deep
convolutional neuralnetworks. In: Machine Learning:Advanced Techniques and
EmergingApplications. 2018. p. 159
17. Goodfellow I, Bengio Y,Courville A. Deep Learning. MIT Press;2016
18. Deng L, Hinton G, Kingsbury B.New types of deep neural networklearning for speech
recognition andrelated applications: An overview. In: 2013 IEEE International
Conference on Acoustics, Speech and Signal Processing.IEEE; 2013. pp. 8599-8603
19. Kriegeskorte N. Deep neuralnetworks: A new framework formodeling biological vision
and braininformation processing. Annual Reviewof Vision Science. 2015;1:417-446
20. Ren S, He K, Girshick R, Sun J. Fasterr-cnn: Towards real-time object detection with
region proposal networks.In: Advances in Neural InformationProcessing Systems.
2015. pp. 91-99

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 11
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

21. Girshick R, Donahue J, Darrell T,Malik J. Rich feature hierarchies foraccurate object
detection and semanticsegmentation. In: Proceedings of theIEEE Conference on
Computer Visionand Pattern Recognition. 2014.p. 580587
22. Girshick R. Fast r-cnn. In:Proceedings of the IEEE InternationalConference on
Computer Vision. 2015.pp. 1440-1448
23. Chorowski JK, Bahdanau D,Serdyuk D, Cho K, Bengio Y. Attentionbasedmodels for
speech recognition. In:Advances in Neural InformationProcessing Systems. 2015. pp.
577-585
24. Kinjal A Joshi, Darshak G. Thakore, “ASurvey on Moving Object Detection
AndTracking in Video Surveillance System,”International Journal of Soft Computing
andEngineering (IJSCE) ISSN: 2231-2307,Volume-2, Issue-3, July 2012.
25. Himani S. Parekh, Darshak G. Thakore,Udesang K. Jaliya, “ A Survey on
ObjectDetection and Tracking Methods,”International Journal of Innovative Researchin
Computer and CommunicationEngineering, Vol. 2, Issue 2, February 2014.
26. Redmon J, Farhadi A. Yolov3: AnIncremental Improvement. arXiv; 2018
27. Redmon J, Divvala S, Girshick R,Farhadi A. You only look once: Unified,real-time
object detection. In:Proceedings of the IEEE Conference onComputer Vision and
PatternRecognition. 2016. pp. 779-788
28. Swati Thorat, Manoj Nagmode, “Detectionand Tracking of Moving
Objects,”International Journal of Innovative Researchin Advanced Engineering
(IJIRAE), Volume1, Issue 1 (April 2014).
29. Everingham M et al. The Pascalvisual object classes (VOC) challenge.International
Journal of ComputerVision. 2010;88(2):303-338
30. Peixoto HM, Teles RS, Luiz JVA,Henriques-Alves AM, Santa Cruz RM.Mice Tracking
Using the YOLOAlgorithm. Vol. 7. PeerJ Preprints; 2019.p. e27880v1
31. Henriques-Alves AM, Queiroz CM.Ethological evaluation of the effects ofsocial defeat
stress in mice: Beyond thesocial interaction ratio. Frontiers inBehavioral Neuroscience.
2016;9:364
32. Jhuang H et al. Automated homecagebehavioural phenotyping of mice.Nature
Communications. 2010;1:68
33. Burgos-Artizzu XP, Dollár P, Lin D,Anderson DJ, Perona P. Social behaviorrecognition
in continuous video. In:2012 IEEE Conference on ComputerVision and Pattern
Recognition. IEEE;2012. pp. 1322-1329
34. Sandeep Singh Sengar, and Susanta Mukhopadhyay. “Foreground Detection via
Background Subtraction and Improved Three-frame Differencing", Arabian Journal for
Science and Engineering, Springer, Vol.-42, Issue-8, pp. 3621–3633, June 2017.
35. Pranjay Shyam, Sandeep Singh Sengar, Kuk-Jin Yoon, and Kyung-Soo Kim. “Robust
Video Enhancement by Adversarial Evaluation of Inter-Frame consistency and
Integrated within Camera-ISP." the 32nd British Machine Vision Conference, 22-25
November, 2021.
36. Pranjay Shyam, Sandeep Singh Sengar, Kuk-Jin Yoon, and Kyung-Soo Kim.
“Exploring Data Efficient Techniques for Image Restoration and Enhancement." In
International Joint Conference on Artificial Intelligence Workshop - Artificial
Intelligence for Autonomous Driving, Montreal, Canada, 21 August 2021.

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 12
Journal of Artificial Intelligence, Machine Learning and Neural Network
ISSN: 2799-1172
Vol: 03, No. 02, Feb-Mar 2023
http://journal.hmjournals.com/index.php/JAIMLNN
DOI: https://doi.org/10.55529/jaimlnn.32.1.13

37. Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu C-Y, et al. SSD: Single shot
multibox detector. In: European Conference on Computer Vision. Cham: Springer;
2016. pp. 21-37
38. Sandeep Singh Sengar. “Motion segmentation based on structure-texture decomposition
and improved three frame differencing", In 15th International Conference on Artificial
Intelligence Applications and Innovations, Crete, Greece, pp. 609–622, May 2019,
Springer.

Copyright The Author(s) 2023.This is an Open Access Article distributed under the CC BY
license. (http://creativecommons.org/licenses/by/4.0/) 13