International Journal of Engineering Applied Sciences and Technology, 2021

Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254

Published Online May 2021 in IJEAST (http://www.ijeast.com)

ARTIFICIAL INTELLIGENCE IN
COMPUTER VISION
Aryan Karn
Motilal Nehru National Institute of Technology Allahabad, Prayagraj
Department of Electronics and Communication Engineering

Abstract- Computer vision is an area of research concerned often employed in visual sign decoding. In addition, it is used in
with assisting computers in seeing. Computer vision issues aim Computer Vision and Natural Language Processing to organize
to infer something about the world from observed picture data material (NLP). It is possible to construct a convolutional neural
at the most abstract level. It is a multidisciplinary subject that network using a variety of structural blocks. These structural
may be loosely classified as a branch of artificial intelligence blocks include convolution layers, pooling layers, and fully
and machine learning, both of which may include using connected layers, all of which will be discussed briefly in this
specific techniques and using general-purpose learning article. In the next sections, the author covers Deep Learning and
methods. As an interdisciplinary field of research, it may seem the many neural network techniques lumped together. In addition,
disorganized, with methods taken and reused from various the book covers Convolutional Neural Networks, their
engineering and computer science disciplines. While one construction, and their applications in several fields, including
specific vision issue may be readily solved with a hand-crafted medicine and engineering.
statistical technique, another may need a vast and
sophisticated ensemble of generic machine learning II. LITERATURE SURVEY
algorithms. Computer vision as a discipline is at the cutting
edge of science. As with any frontier, it is thrilling and chaotic, A. Deep Learning and Neural Networks
with often no trustworthy authority to turn to. Numerous
beneficial concepts lack a theoretical foundation, and some Machine Learning is a subset of Deep Learning, a subset of
theories are rendered ineffective in reality; developed regions Artificial Intelligence (AI). Machine learning uses algorithms and
are widely dispersed, and often one seems totally unreachable training data to automatically detect patterns and with little human
from the other. intervention. Artificial Intelligence is a method for teaching
computers to act like humans. At the same time, Deep Learning is
Keywords—Computer vision, Artificial intelligence, Neural networks, inspired by the structure and function of the human brain, as
CNN., Deep learning, machine learning represented symbolically by an artificial neural network. [12]
While deep learning was originally proposed in the 1980s, it has
I. INTRODUCTION shown significant benefits in recent years for two primary reasons:

Recently, computer vision has gained traction and popularity A. This requires a significant level of knowledge. For instance, the
as a consequence of the many applications it has found in areas development of autonomous vehicles necessitates the collection of
like health and medical, sports and entertainment, automaton many pictures and lengthy video recordings.
design, and self-driving cars. Many of these applications rely on
visual recognition tasks such as image order, restriction, and B. Deep learning requires a large capacity for recording. High-
identification. Recent advances in Convolutional Neural Networks performance GPUs offer an efficient parallel design that is well-
(CNNs) have resulted in an extraordinary performance in these suited for deep learning. When used in conjunction with clusters
best-in-class visual recognition assignments and frameworks, or cloud computing, this significantly lowers the time required to
demonstrating the power of Convolutional Neural Networks. train a deep learning network from weeks to hours or less. [11].
Consequently, convolutional neural networks (CNNs) have Deep learning may be used to solve a wide range of problems. For
emerged as the basic building blocks of deep learning example, the author discusses autonomous driving, aerospace and
computations in computer vision. military, medical research, industrial automation, and electronics
in more detail in the closing part of the article.
Deep Neural Networks (DNN) is a kind of neural network that has
better image identification skills and is often utilized in computer In general, a Neural Network is a kind of algorithm that accepts
vision computations. Convolutional Neural Networks (CNN or certain input parameters and processes them using an Activation
ConvNet) is a subtype of Deep Neural Networks (DNNs) that are Function to get the desired Output. In this method, the input to

249
 International Journal of Engineering Applied Sciences and Technology, 2021
Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254
Published Online May 2021 in IJEAST (http://www.ijeast.com)

output processing component is referred to as a neuron. Consider In deep learning, a convolutional neural network (CNN), often
the fundamental example of calculating the purchase price of a known as a ConvNet, is a kind of deep neural network that is
home. Numerous factors must be taken into account, each of frequently used to analyze visual pictures. In certain areas, it is
which has an impact on the Price component. For instance, the also referred to as a convolutional neural network (CNN). These
square footage of the room, the number of bedrooms, and the zip artificial neural networks are referred to as shift-invariant artificial
code. Thus, if we take price as an Output, the following Neural neural networks or space-invariant artificial neural networks due
Network shows how a neural network could produce that Output to their shared-weights architecture and translation invariance
using the parameters stated earlier as inputs. properties (SIANNs). Algorithms may be used to identify pictures
and videos, create recommender systems, categorize images, do
medical image analysis, and evaluate natural language. In the next
part, the author discusses what Convolution is, how it extracts data
from pictures, and the architecture and components of CNN,
among other things. This will show how CNN examines the
content of an image and processes the data in order to provide the
intended result to the audience.

C. Architectural Overview

Fig 1: An example of Standard Neural Network [1] Convolution is a mathematical procedure that takes two functions
and produces a third function that illustrates how the shape of one
Each circle represents a neuron that is given an Activation function is affected by the shape of the other. To complete the
Function that computes the desired Output by combining distinct operation, the Convolution process requires the calculation of the
values for various input parameters. The Activation Function is Result function, as well as the initialization of the Result function.
determined by the algorithm's purpose/application. For instance, Convolution is a data processing technique that entails
each circle represents a neuron that is given an Activation categorizing the components (content) of an image in order to
Function that computes the desired Output by combining distinct assist Machine Learning and ultimately generate the desired
values for various input parameters. The Activation Function is Output through the algorithm. It is utilized in the processing of
determined by the algorithm's purpose/application. For instance, picture data. Deep Learning and Neural Networks are two
in the above example, the objective is to determine the maximum different types of neural networks that are capable of analyzing
price of a house. For the sake of simplicity, let us suppose that the image data. Deep Learning is a kind of neural network that enables
Output is solely dependent on two input variables, namely the size data-driven learning. As indicated by the procedure's name, the
and number of bedrooms. In this instance, the bigger the house and convolution process separates the wheat from the chaff.
the more bedrooms, the greater the price of the house. Thus, the
Activation Function (Neuron) will be defined in such a manner This structure may be seen as a three-dimensional volume of
that it will pick the greatest possible value for each input parameter neurons in a cellular environment. A distinguishing feature of how
and then compute the Output. Obviously, this seems to be very CNNs have evolved from earlier feed-forward versions is their
easy in this case, but when a large number of factors are involved, ability to improve computational efficiency via the addition of
decision-making is not as straightforward as it appears based on new layer types to their design. How about we take a closer look
maximum or minimum values alone. And here is where Data- at the general design of CNNs right now? [4]
Driven Machine Learning comes into play. The method takes
advantage of data saved (learned!) from previous instances in D. Basic CNN components
order to calculate the optimal Output using the Activation
Function. The above example shows a Standard Neural Network, 1. Convolutional Layer:
which is often used to generate Output from statistical, numerical,
CNN, or convolutional neural network, is a kind of neural
and other quantitative data. The kind of Neural Network to employ
network model that is designed for dealing with two-dimensional
is determined by the nature of the input data that the algorithm
image data, although it may also be used to deal with one-
must handle. The following table summarizes the capabilities of
dimensional and three-dimensional data. Convolution is
different Neural Networks in processing various kinds of input
accomplished via the use of a channel (a small matrix whose size
data. [1] In the remainder of this article, the author will concentrate
may be chosen). In this channel, which travels the whole picture
only on the Convolutional Neural Network method used in Deep
network, the task is to reproduce the image's features by utilizing
Learning.
the pixel values that were first used. Each of these increases is
B. Deep Learning using Convolutional Neural Network for added together to form a single number towards the end of the
Computer Vision process. When doing a comparison action, the channel moves

250
 International Journal of Engineering Applied Sciences and Technology, 2021
Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254
Published Online May 2021 in IJEAST (http://www.ijeast.com)

rightward by n units (this number may vary). After traversing over

each place, a framework is produced that is much less in size than
the information grid that was previously constructed. [7]

Fig 6: Matrix calculation in convolutional layer [14]

Fig 3: Architecture of a CNN [6]

Edge Detection Example:

In Figures 4 and 5 below, to detect the horizontal and vertical

images with the help of a matrix, let's consider a greyscale image,
with a 6 x 6 matrix and a filter of 3x3 applied to it.[14]

Fig 7: Convolving over the entire image [14]

The way to detect the vertical edge in the image is to look for the
pixel values as, if the pixel values are greater, then brightness at
that part of the image will be more, and if the value is less, it will
be dark. [14]

2. Pooling Layer:

Spatial pooling (alternatively referred to as subsampling or

Fig 4: Identifying the edges [14] down sampling) lowers the dimensionality of each element map
while preserving the most critical data. Spatial pooling may occur
in a number of ways: Quantifiers include the terms maximum,
average, and total. If Max Pooling occurs, we define a spatial
neighborhood (for example, a 22-window neighborhood) and
choose the biggest component from the redressed highlight map
contained inside that neighborhood. Rather than choosing the
biggest component, we might choose the average (Average
Pooling) or a total of all components included inside that window.
Max Pooling has been shown to be increasingly effective with
time. [8] Max pooling, as shown below, chooses the component
Fig 5: Scanning of the image pattern [14] with the largest size from the rectified feature map. Choosing the
biggest component is equal to using the conventional pooling
After the above calculations of the matrix, we get the matrix as method. The phrase "sum pooling" refers to the gathering of all
shown in fig:6. To calculate it, we take the initial 3 X 3 framework components in an element map. [8]
from the 6 X 6 picture and increase it with the channel. Let’s
consider the following matrix of 4x4 order and the calculation
takes place as: for example, 3*1 + 0 + 1*-1 + 1*1 + 5*0 + 8*-1 +
2*1 + 7*0 + 2*-1 = - 5. To compute the second component of the
4 X 4 order, we will move the channel one step ahead to the right
side of the original Greyscale matrix and so on: [14]

251
 International Journal of Engineering Applied Sciences and Technology, 2021
Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254
Published Online May 2021 in IJEAST (http://www.ijeast.com)

Parkinson's illness, malignant uterine growth, and a host of other

medical issues, and the technology is getting more advanced. With
the use of best-in-class image processing technology and
computer vision methods, early identification of any potential
diseases will be feasible. In this manner, treatment may be
administered at an inconvenient time during the disease or, in any
event, the likelihood of their recurring is decreased [2].

Fig 8: Max Pooling [8]

3. Fully Connected Layers:

Fully Linked layers have every neuron in the layer above it

connected to every neuron in the layer below it. To put it simply, Fig 10: Cross-section of the 3D image of CT Scan and MRI [2]
FC works in the same manner as a conventional neural network,
such as a Multi-Layer Perceptron, does (MLP). The main
distinction is that information sources would be molded and
organized in the manner defined by earlier phases of a CNN, rather B. Automobile:
than the other way around. [7]. As illustrated in the diagram below,
With the expanded publicity of oneself driving autos, car
the feature map matrix is converted into a vector as(x1,x2,...xn) by
businesses are vigorously subject to Computer Vision since it is
utilizing the FC layer, and the resulting vectors are merged to
intended for understanding the driving condition, including
create a model. Then, using the activation function, we can
identifying impediments, people on footpaths, and conceivable
classify the Output into different categories.
crash ways. Self-driving autos are gradually advancing into the
market, with more organizations searching for imaginative
approaches to bring progressively electric vehicles onto the street.
Computer Vision innovation enables these self-driving vehicles 'to
see' the earth while AI calculations make the "minds" that help that
Computer Vision translate the items around the vehicle. Self-
driving autos are furnished with numerous cameras to give a total
360-degree perspective on nature inside the scope of several
meters. Tesla vehicles, for example, utilize something like 8
encompasses cameras to accomplish this accomplishment. Twelve
ultrasonic sensors for identifying hard and delicate articles out and
about and a front-oriented radar that empowers the identification
of different vehicles even through downpour or mist are
additionally introduced to supplement the cameras. With a lot of
information being encouraged into the vehicle, a basic PC won't
Fig 9: Fully connected Layer [7] be sufficient to deal with the inundation of data. This is the reason
all self-driving autos have a locally available PC with Computer
III APPLICATIONS Vision highlights made through AI. The cameras and sensors are
entrusted to both recognize and group protests in nature - like
A. HealthCare:
people on foot. The area, thickness, shape, and profundity of the
Computer vision is extensively employed in the diagnosis of items must be considered quickly to empower the remainder of the
diseases by analyzing X-rays, magnetic resonance imaging (MRI), driving framework to settle on proper choices. Every one of these
and other medical pictures. It has been shown to be just as calculations is just conceivable through the incorporation of AI
convincing as traditional human specialists in the area when it and deep neural systems, which results in highlights like the
comes to accuracy. On a regular basis, Computer Vision is person on foot recognition [15].
effectively diagnosing pneumonia, cerebrum tumor’s, diabetes,

252
 International Journal of Engineering Applied Sciences and Technology, 2021
Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254
Published Online May 2021 in IJEAST (http://www.ijeast.com)

increasingly fascinating as an ever-increasing quantity of

information and processing power became accessible.

(2).AlexNet (2012) – In 2012, Alex Krizhevsky (together with

others) published AlexNet, which was a more in-depth and much
more complete version of the LeNet. AlexNet was the clear winner
of the inaugural ImageNet Large Scale Visual Recognition
Challenge (ILSVRC) in 2012, outperforming the competition by
a wide margin. This research represented a major advancement
over prior techniques, and the widespread use of CNNs today may
be linked back to the findings of this study.

A Convolutional Network created by Matthew Zeiler and Rob

Fig 11: Tesla Car’s Vision, Source: Tesla [15] Fergus was presented at the ILSVRC 2013 as part of the 3.ZF Net
(2013) session. The ZDNet was the moniker that was given to this
C. Astronomy: network (short for Zeiler and Fergus Net). It was feasible to make
improvements to AlexNet by altering the design hyperparameters
Our full understanding of the universe is based on photon used in its creation. When Szegedy and colleagues from Google
estimations, which are mostly composed of pictures of the presented a Convolutional Network at the ILSVRC 2014
universe. This opens the door to the potential of utilizing conference, it was given the moniker "GoogleLeNet" (2014). This
Computer Vision in astronomy since our universe is so enormous, organization's main goal was the creation of an Inception Module
and our universe's lone natural rule predicts that the data gathered that significantly decreased the number of parameters in the
will be just as large. It will be impossible for the stargazer, or for system (4M, contrasted with AlexNet with 60M).
anybody else, to physically contemplate this information in its
entirety. We can decipher all of the data in a short period of time, In the 2014 International Laser Scanning and Vision Research
thanks to Computer Vision. To put it another way, computer vision Conference (ILSVRC), a system that became known as the
is currently being utilized to find new planets and big bodies, with VGGNet was the first to cross the finish line. In particular, it aimed
applications such as exoplanet imaging, star and cosmic system at demonstrating how important it is for efficient execution to have
grouping, and other similar tasks [3]. a system with sufficient depth (i.e., layers). It was ResNets (2015),
a Residual Network created by Kaiming He (and others), that was
D. Industrial: awarded sixth place in the ILSVRC 2015. ResNets (2015) was the
winner of the ILSVRC 2015. Convolutional Neural Network
In Industries, Computer Vision is utilized on the mechanical models such as ResNets are currently by a wide margin the best-
production systems for checking groups, identifying harmed parts, in-class models, and they will continue to be the default option for
for the examination of the completed merchandise. Here, Machine utilizing ConvNets for the foreseeable future (as of May 2016).
Vision apparatuses help in discovering infinitesimal level
surrenders in items that basically can't be distinguished through The seventh source is DenseNet, which was launched in August
human vision. In assembling undertakings, perusing scanner tags 2016. A network of nodes that are closely packed together. This
or QR code are fundamental as they give a one of a kind densely linked convolutional network, developed by Gao Huang
recognizable proof to an item. Perusing a great many standardized (and others) and published recently, has each layer directly
identifications in a day isn't a simple errand for people; at the same connected to every other layer in a feed-forward architecture, with
time, it very well may be done effectively in minutes through each layer being straightforwardly correlated with each other
Computer Vision [3]. layer. Following the completion of five highly concentrated article
acknowledgment benchmark assignments, the DenseNet was
IV ConvNet ARCHITECTURE found to have gained substantial gains over prior best-in-class
architectures, results revealed. View this video to see exactly how
Convolutional Neural Networks (CNNs) is a kind of neural the Torch was carried out.
network that has been around since the mid-90s. You'll discover
some more visually arresting designs in the section below [9]. V. CONCLUSION

A convolutional neural network was in the process of being At the beginning of the paper, we discussed the overview of
created from the late 1990s to the middle of the 2010s, and it was deep learning and how Neural networks in dep learning are
known as LeNet during that time period. The tasks that deployed to process various inputs to gain desired outputs. In the
convolutional neural networks were capable of doing grew later part, the author has focused on Convolutional Neural

253
 International Journal of Engineering Applied Sciences and Technology, 2021
Vol. 6, Issue 1, ISSN No. 2455-2143, Pages 249-254
Published Online May 2021 in IJEAST (http://www.ijeast.com)

Network and explained in detail a convolution operation, the medium.com/alumnaiacademy/introduction-to-computer-vision-

system architecture of CNN, and how the layers of CNN work in 4fc2a2ba9dc.
coordination to identify the highlights and the patterns of an
image. Using these algorithms author has described how CNN can [9] Ujjwalkarn. “An Intuitive Explanation of Convolutional
be applied in various industries. Through this paper, it can be Neural Networks.” The Data Science Blog, 29 May 2017,
concluded that CNN has become a very powerful tool in machine https://www.ujjwalkarn.me/2016/08/11/intuitive-explanation-
learning. By providing various images as input data at the machine convnets/.
learning phase can facilitate the learning process faster, and the
data can be deployed for multiple-output functions, which is a [10] Gibson, Adam, and Josh Patterson. “Deep Learning.”
major advantage of CNN. Apart from the application, the author O'Reilly | Safari, O'Reilly Media, Inc.,
mentioned in the earlier section that CNN is now also being
considered for IoT, Commercial, and domestic security systems. www.oreilly.com/library/view/deep-
Thus, CNN has gained a very prominent place in Data Engineering learning/9781491924570/ch04.html
and still is gaining.
[11] “What Is Deep Learning?: How It Works, Techniques &
VI REFERENCES Applications.” How It Works, Techniques & Applications -
MATLAB & Simulink,
[1] Pulkit Sharma, Analytics Vidhya, An Introductory Guide to www.mathworks.com/discovery/deep-learning.html.
Deep Learning and Neural Networks, 22 Oct. 2018,
https://www.analyticsvidhya.com/blog/2018/10/introduction- [12] Brownlee, Jason. “What Is Deep Learning?” Machine
neural-networks-deep-learning/ Learning Mastery, 31 Oct. 2019,
machinelearningmastery.com/what-is-deep-learning/.
[2] Khan, S. A Guide to Convolutional Neural Networks for
Computer Vision. Morgan & Claypool, 2018.

[3] Verma, Shiva. “Understanding 1D and 3D Convolution Neural

Network: Keras.” Medium, Towards Data Science, 1 Oct. 2019,

towardsdatascience.com/understanding-1d-and-3d-convolution-
neural-network-keras-9d8f76e29610.

[4] Upadhyay, Yash. “Computer Vision: A Study on Different

CNN Architectures and Their Applications.” Medium, AlumnAI
Academy, 29 Mar. 2019,
medium.com/alumnaiacademy/introduction-to-computer-vision-
4fc2a2ba9dc

[5]CS231n Convolutional Neural Networks for Visual

Recognition, cs231n.github.io/convolutional-networks/.

[6] “Autopilot.” Tesla, Inc, www.tesla.com/autopilot.

[7] Deshpande, Adit. “A Beginner's Guide To Understanding

Convolutional Neural Networks.”

A Beginner's Guide To Understanding Convolutional Neural

Networks – Adit Deshpande – Engineering at forwarding | UCLA
CS '19,

adeshpande3.github.io/adeshpande3.github.io/A-Beginner's-
Guide-To-Understanding-Convolutional-Neural-Networks/.

[8] Upadhyay, Yash. “Computer Vision: A Study On Different

CNN Architectures and Their Applications.” Medium, AlumnAI
Academy, 29 Mar. 2019,

254