A Project Report

On

“UNIFORM AND ATTENDANCE DETECTION USING DEEP

LEARNING”

Submitted for partial fulfillment of award of the degree

BACHELOR OF TECHNOLOGY

in
“COMPUTER SCIENCE & ENGINEERING”

by

Dheeraj Yadav (1901200100122)

Priyanshu Rai (1901200100122)
Riya Kesharwani (1901200100122)
Vandana Singh (1901200100122)

Under Guidance of
Mr. Ajay Gupta
(Assistant Professor, Department of CSE)

DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

INSTITUTE OF TECHNOLOGY & MANAGEMENT, GIDA, GORAKHPUR
(AFFILLIATED TO DR. A.P.J. ABDUL KALAM TECHNICAL UNIVERSITY, LUCKNOW)

SESSION: 2024-25
 INSTITUTE OF TECHNOLOGY & MANAGEMENT
Integrated Technical Campus: Engineering, Pharmacy & Management
Approved by AICTE, Pharmacy Council of India, New Delhi & Affiliated to Dr. APJAKTU, Lucknow
AL-1, Sector - 7, GIDA, Gorakhpur - 273209 (UP)

Department of Computer Science & Engineering

Vision & Mission of the Department

Vision of the Department:

Excellence in the education to produce competent computer professional to meet

need of industry and society.

Mission of the Department:

1- To impart quality education in computer science & engineering through

competent faculties, good infrastructure and learning resources to produce
competent graduates required by the industry.

2- To inculcate knowledge and skill sets and to strengthen industry-institute

interaction to promote individual and team excellence to meet the requirement of
industry.

3- To develop graduates to be computer professionals with high ethical standard

and self-learning ability.

Program Educational Objective

PEO No. Program Educational Objectives Statements

To prepare competent computer science graduates to have successful
PEO1
career in application software development and in allied industries.
To impart in depth knowledge in core computer science and engineering
PEO2 domain and expertise in providing solution to complex problems using
modern IT tools to meet the requirement of industries.
To inculcate professional skills and to work effectively as individual and
PEO3
member of a team in multidisciplinary projects.
To develop self learning ability, to protect environment and exercise
PEO4
ethical code of conduct in professional practice and activities.

ii | P a g e
 INSTITUTE OF TECHNOLOGY & MANAGEMENT
Integrated Technical Campus: Engineering, Pharmacy & Management
Approved by AICTE, Pharmacy Council of India, New Delhi & Affiliated to Dr. APJAKTU, Lucknow
AL-1, Sector - 7, GIDA, Gorakhpur - 273209 (UP)

Department of Computer Science & Engineering

Rubrics 1 is tagged with course outcome to evaluate student’s performance in

Zeroth review.

Rubrics Course outcome PO/PSO mapping Skill Evaluated

Able to identify the project and review
research literature relevant to the
PO1, PO2 & PO12/
1 problem and able to apply recent Literature Survey
PSO1 & PSO2
trends and technologies in developing
project.

Rubrics 2, 3, 4, 5 and 6 are tagged with course outcome to evaluate student’s

performance in 2nd and 3rd review.

Rubric
Course outcome PO/PSO mapping Skill Evaluated
s
Able to classify the projects and
considering the impacts of safety,
PO6, PO7 &
2 environment and sustainability, Classification
PO8/PSO1 & PSO2
society, ethics and cost while planning
and during execution of the project.
Apply research-based knowledge to
design and develop the project using
appropriate software programming Modern IT Tools
3 PO3 & PO5
skills to meet the specification with Uses
consideration for cost, cultural, social
and environmental.
Demonstrate to function effectively as
Individual and team
individual and as a member in the PO9 & PO11/PSO1 &
4 work & project
group to contribute and manage the PSO2
management
project.
Apply written and verbal
communication through report and PO10 & PO11/ PSO1
5 Communication
presentation to monitor and evaluate & PSO2
the project.
Able to identify and formulate real life Modern IT Tools
6 PO2, PO5 & PO12
problem into project. Uses

iii | P a g e
 INSTITUTE OF TECHNOLOGY & MANAGEMENT
Integrated Technical Campus: Engineering, Pharmacy & Management
Approved by AICTE, Pharmacy Council of India, New Delhi & Affiliated to Dr. APJAKTU, Lucknow
AL-1, Sector - 7, GIDA, Gorakhpur - 273209 (UP)

Department of Computer Science & Engineering

Program Name : B. Tech (CSE) Semester : VIII

Course Name : Project Session : 2024-25
Course Code : Section :

Continuous Internal Evaluation (CIE): Phase-0

Sr. No. Roll No. Name CO1
Rubrics Rub 1 Total
Maximum Marks
1.
2.
3.
4.

Continuous Internal Evaluation (CIE): Phase-I

Sr. No. Roll No. Name CO2 CO3 CO4 CO5 CO6
Total
Rubrics Rub 2 Rub 3 Rub 4 Rub 5 Rub 6
Maximum Marks
1.
2.
3.
4.
Continuous Internal Evaluation (CIE): Phase-II
Sr. No. Roll No. Name CO2 CO3 CO4 CO5 CO6
Total
Rubrics Rub 2 Rub 3 Rub 4 Rub 5 Rub 6
Maximum Marks
1.
2.
3.
4.

Grand Totals of all Phases:

Sr. No. Roll No. Name Phase-0 Phase-I Phase-II Total
Maximum Marks
1.
2.
3.
4.

iv | P a g e
 INSTITUTE OF TECHNOLOGY & MANAGEMENT
Integrated Technical Campus: Engineering, Pharmacy & Management
Approved by AICTE, Pharmacy Council of India, New Delhi & Affiliated to Dr. APJAKTU, Lucknow
AL-1, Sector - 7, GIDA, Gorakhpur - 273209 (UP)

Department of Computer Science & Engineering

Signature

v|Page
 DECLARATION

I hereby declare that the project work entitled “Plant Leaf Disease Detection Model”,

has been prepared by us during the year 2024-25 under the guidance of “Mr. Ajay

Gupta”, Assistant Professor, Department Of Computer Science & Engineering, Institute

of Technology and Management, and this project work is submitted in the partial

fulfillment of the requirements for the award of the degree of Bachelor of Technology in

Computer Science & Engineering.

I also declare that this project is the outcome of our own efforts that it has not

been submitted to any other University or Institute for the award of any degree.

Name (Roll Number) Sign Here

Name (Roll Number)

Name (Roll Number)

Name (Roll Number)

Place : Gorakhpur
Date :

v|Page
 CERTIFICATE

This is to certify that the project report entitled Plant Leaf Disease Detection Model

Using Convolutional Neural Network (CNN) submitted by Nidhi Chaudhary,

PoojaYadav, Preeti Kumari, Deeksha Srivastava to the Institute of Technology and

Management GIDA, Gorakhpur, in partial fulfillment for the award of the degree of B.

Tech in (Computer Science and Engineering) is a bona fide record of project work

carried out by him/her under my/our supervision. The contents of this report, in full or in

parts, have not been submitted to any other Institution or University for the award of any

degree.

(Signature)
Name of your guide/supervisor
Assistant Professor
Department of Computer Science and Engineering
Institute of Technology and Management, GIDA, Gorakhpur, UP,
Pin- 273209

Date :

vi | P a g e
 ACKNOWLEDGEMENT

I would like to thank my project guide “Mr. Ajay Gupta”, Assistant Professor,
Department of Computer Science & Engineering, Institute of Technology and
Management, GIDA, Gorakhpur, U.P. for his valuable guidance and suggestions. I
would like to thank my project coordinator “Mr. Nitin Dixit”, Associate Professor,
Department of Computer Science & Engineering, Institute of Technology and
Management, GIDA, Gorakhpur, U.P. for his valuable guidance and suggestions. I am
thankful for his/her continual encouragement, support, and invaluable suggestions.
Without his encouragement and guidance, this project would not have been materialized.
Throughout the writing of the project, I have received a great deal of support and
assistance.

I am very thankful to HOD “Mr. Ashutosh Kumar Rao” Department of Computer

Science and Engineering, for his kind cooperation. I would also like to thank to the
Honorable Director “Dr. N. K. Singh” for his kind help and support. I would also like to
thanks to all my friends who continuously supported me.

I want to express my appreciation to every person who contributed either with

inspirational or actual work to this project. Finally, I must express my very profound
gratitude to my parents. Thanks to all.

Nidhi Chaudhary (2001200109010)

Pooja Yadav (2001200109008)
Preeti Kumari (2001200109011)
Deeksha Srivastava(2001200109006)

vii | P a g e
 ABSTRACT

As India's population grows and there is a greater need for food, plant diseases become a

serious threat to agricultural output and have a serious impact on farmers. Plant diseases

can be anonymously detected, helping to ensure food preservation and limit Monetary

Impairment. Images of sick plants can be used to illustrate the illnesses. Convolutional

neural networks (CNN) are used for their assortment capabilities to produce genuine

output. These publications conclude a graph on additional categorization strategies for

plant leaf diseases assortment. For farmers, spotting symptoms of illness with the naked

eye is difficult. Crop defense in a big frame Using digital image processing technology

that can identify sickness, the work is done. Diseases can significantly diminish crop

origin, which poses a serious threat to food security.

Accurately identifying plant diseases is so crucial and essential. Convolutional neural

networks (CNN), which are frequently used to categorize plant diseases, are the

foundation of this system. They represent traditional categorization techniques and have

either fully or partially addressed the problems with contemporary technology in this

area. In this study, we looked at the most current CNN networks that were pertinent to

categorizing plant leaf diseases. Plant diseases are one of the main drivers for boosting

food production and lowering production-related harm. Rapid diagnosis and detection of

crop diseases are inevitable. Recently created deep learning algorithms have shown to be

helpful in recognizing plant diseases, giving a practical tool with incredible accuracy.

Keywords: Detection, Recognition

viii | P a g e
ix | P a g e
 TABLE OF CONTENTS

Page No.
CERTIFICATE v
DECLARATION vi
ABSTRACT vii
ACKNOWLEDGEMENT viii
TABLE OF CONTENTS ix
LIST OF TABLES x
LIST OF FIGURES xi
LIST OF SYMBOLS AND ABBREVIATIONS xii
CHAPTER 1 : INTRODUCTION 1-68
1.1 GENERAL 1
1.2 INTRODUCTION TO MICROWAVE SOURCES 2
1.3 GENERAL 3
1.3.1 Evolution of Microwave Tubes 5
1.3.2 Operating Principle of Gyro-TWT 6
1.3.2.1 General 15
1.3.2.1 Beam-mode line dispersion 16
1.3.2.3 Waveguide mode dispersion 18
CHAPTER 2 : LITERATURE REVIEW 59-150
2.1 75

x|Page
 LIST OF FIGURES

Figure No. Title Page No.

Figure 1.1 Block Diagram of the General Framework 3

xi | P a g e
 LIST OF TABLES

Table No. Title Page No.

Table 1.1 Block Diagram of the General Framework 5

xii | P a g e
 LIST OF ABBREVIATIONS

ABBREVIATIONS NAME
CNN : Convolutional Neural Network
:
:
:
:
:

xiii | P a g e
CHAPTER 1
INTRODUCTION

In this chapter we learn about most modern deep learning models are based on artificial
neural networks, specifically convolutional neural networks (CNN)s, although they
can also include propositional formulas or latent variables organized layer-wise in
deep generative models such as the nodes in deep belief networks.

Artificial Intelligence
Since the beginning of the introduction of technology, humans have had an obsession
with automation. AI enables machines to think independently of human input. The field
of computer science is very vast. Three such kind AI systems exist: ANI, or artificial
narrow intelligence, which is task and goal-focused. Artificial General Intelligence
(AGI), which enables robots to learn, comprehend, and behave in a manner that is
identical to that of humans in a certain circumstance. In a hypothetical AI system
known as ASI (Artificial Super Intelligence), machines might be able to explain
intelligence that is greater than that of the smartest humans [1].

Artificial Intelligence

Artificial Neural
Network

Deep Learning

Fig1.1 Subset of Artificial intelligence

1|Page
 Artificial Neural Networks
Artificial neural networks (ANNs), usually simply called neural networks (NNs) or
neural nets, are computing systems inspired by the biological neural networks that
constitute animal brains.

An ANN is based on a collection of connected units or nodes called artificial neurons,

which loosely model the neurons in a biological brain. Each connection, like the
synapses in a biological brain, can transmit a signal to other neurons. An artificial
neuron receives signals then processes them and can signal neurons connected to it. The
"signal" at a connection is a real number, and the output of each neuron is computed by
some non-linear function of the sum of its inputs. The connections are called edges.
Neurons and edges typically have a weight that adjusts as learning proceeds. The weight
increases or decreases the strength of the signal at a connection. Neurons may have a
threshold such that a signal is sent only if the aggregate signal crosses that threshold
[10].

Neural networks
A neural network is a set of algorithms that mimics how the human brain works in order
to find hidden connections in a piece of data.

Fig1.2 Neural Network

When you open your eyes, what you see is typically described as "data," which is
processed by

2|Page
the neurons (data processing cells) in your brain and used to recognise what is around
you. These parallels in neural network operation demonstrate how they analyze a huge
data set, find the patterns, and then output the findings [6]. There are some types of
neural networks.

Types of Neural Network

There are some types of neural network as given below
i. Perceptron
ii. Multi-Layer Perceptron
iii. Radial Basis Networks
iv. Feed Forward Network
v. Long Short-Term Memory Networks
Perceptron
The simplest and most traditional type of neural network is the perceptron. It consists of
just one neuron, which takes in input and uses an activation function to transform it into
a binary output. The only jobs it may be used for are binary classification tasks, and it
has no hidden layers. The processing of the addition of input values with their weights
is handled by the neuron. The activation function receives the resultant sum and outputs
a binary value.

Fig 1.3 Perceptron

3|Page
 Multi-Layer Perceptron
The fundamental drawback of feed-forward networks was that they couldn't learn
by back propagation. Neural networks with many hidden layers and activation
mechanisms are called multi-layer Perceptron. The learning is done under supervision,
and gradient descent is used to

Fig 1.4 Multi-Layer Perceptron

update the weights. The multi-layer perceptron is bi-directional, meaning that the inputs
are transmitted forward and the weight updates are transmitted backward. With regard
to the type of target, the activation functions can be altered. The most common
classification methods are Softmax for multi-class classification, Sigmoid for binary
classification, and so on. Because every neuron in one layer is connected to every
neuron in the following layer, these are also known as "dense networks."

Radial Basis Networks

The targets are predicted entirely differently by Radial Basis Networks (RBN). There
are three layers to it: an input layer, an RBF neuron layer, and an output layer. For each
instance of the training data, the RBF neurons keep track of the real classes. The radial
function utilized as an activation function distinguishes the RBN from the typical
multilayer perceptron

4|Page
 Fig 1.5 Radial Basis Networks

Feed Forward Network

The hidden layers and numerous neurons that make up Feed Forward (FF) networks are
interconnected. Since there is no backward propagation and just forward data flow,
these are referred to as "feed-forward" systems. Depending on the application, the
network may not always have hidden levels

Fig 1.6 Feed Forward Networks

5|Page
Feed Forward networks are used in:
i. Classification
ii. Speech recognition
iii. Face recognition
iv. Pattern recognition

Long Short-Term Memory Networks

The fading gradient problem in RNNs is resolved by LSTM neural networks by
including a unique memory cell that can retain data for a very long time. To specify
which outputs should be utilized or disregarded, LSTM employs gates. An input gate,
an output gate, and a forget gate are the three gates that are utilized. What information
should be stored in memory is decided by the input gate. The output gate regulates the
data sent to the following layer, while the forget gate regulates when to discard or forget
the unnecessary data [2].

LSTMs are used in various applications such as

i. Gesture recognition
ii. Speech recognition
iii. Text prediction

As with ANNs, many issues can arise with naively trained DNNs. Two common issues
are over- fitting and computation time in ANN. DNNs are prone to over fitting because
of the added layers of abstraction, which allow them to model rare dependencies in the
training data. Deep Neural Network is able to solve far more complex problems through
a wide range of architectures other than simple feed-forward, fully connected networks.

Deep Neural Network

It is a network with more than two layers and the word “deep” refers to the number of
layers through which the data is transformed. A deep neural network (DNN) is an
artificial neural network (ANN) with multiple layers between the input and output
layers. There are different types of neural networks but they always consist of the same
components: neurons, synapses, weights, biases, and functions. These components as a
whole function similarly to a human brain, and can be trained like any other ML
algorithm. DNNs can model complex non-linear relationships. DNN architectures
generate compositional models where the object is expressed as a layered composition

6|Page
of primitives. The extra layers enable composition of features from lower layers,
potentially modeling complex data with fewer units than a similarly performing shallow
network.

There are many different ways of designing a DNN, using different layering structures,
different types of layers, and different ways of connecting the nodes [4].

Types of Deep Neural Network

Following types of deep neural networks are popularly used today [2]
i. Convolutional Neural Networks (CNN)
ii. Recurrent Neural Networks (RNN)
iii. Deep belief network (DBN)
iv. Restricted Boltzmann Networks or Autoencoders (RBNs)

Convolutional Neural Networks

The most popular neural networks for image categorization are convolutional neural
networks (CNN). Multiple convolution layers of CNN are in charge of extracting
significant features from the image. Low-level details are handled by the earlier layers,
and more high-level aspects are handled by the later levels. Back propagation is used to
update these filters after their initialization at random.

History of Convolutional Neural Networks

With its live coverage of the Challenger space shuttle catastrophe in 1986, the network
defeated the competition. With its live "in-country" telecasts of the Persian Gulf
War five years later, CNN once again outperformed the other networks. CNN's team of
reporters, which included Bernard Shaw, Peter Arnett, and John Holliman, covered the
battle from both sides of the conflict. Daniel Schorr, Wolf Blitzer, Catherine Crier,
Mary Alice Williams, Christiane Amanpour, and Paula Zahn are a few other well-
known CNN reporters and pundits. James Earl Jones, a renowned actor, serves as the
"voice of CNN," announcing in a recorded voice, "This is CNN."A convolutional neural
network, or CNN, is a deep learning neural network sketched for processing structured
arrays of data such as portrayals.

7|Page
CNN, also known as a convolutional neural network, is a deep learning discipline that is
mostly utilized for the processing of visual data. A class of deep feed-forward artificial
neural networks includes CNNs (ANN). For training, this neural network makes
predictions about potential future label assignments using the dataset that has already
been provided. In order to overcome the dimensionality curse, this neural network
makes use of its advantages. Image identification, categorization, captioning, object
detection, and other applications are a few of the areas where CNNs are widely used.
When Alex found the CNNs in 2012, they quickly became quite well- known.

The developers have developed it to the point where an older, 8-layer Alex Net may
now be transformed into a 152-layer ResNet in just three years. CNNs are useful for
jobs that take recommendation systems, contextual relevance, or natural language
processing (NLP) into account. Making sure the neural network processes all the layers
and, consequently, automatically recognizes all the underlying features, is its primary
responsibility

Fig 1.7 History of CNN

A CNN is a convolutional tool that divides the various visual highlights for
analysis and forecasting. The fundamental benefit of CNNs is that they offer a dense,

8|Page
effective network for prediction, identification, and other tasks. In the very large field of
deep learning, CNNs are by far the most discussed topic, and this is typically due to the
CNN.

Huge datasets are used with CNNs; in fact, it is thought that the more data used, the
more accurate the model will be; otherwise, additional processes, like transfer learning,
would be used to increase the dataset. CNN's strength is in its self-sufficient ability to
identify distinctive elements in photos without actual human participation. [7]

The pooling layer, which comes after the convolution layer, is in charge of combining
the maps created by the convolutional layer. CNNs also offer the option of adding
dropout layers for regularization, which disable or turn off specific neurons to lessen
over fitting and accelerate convergence. Rectified Linear Unit (ReLu) activation
functions are employed in CNNs' hidden layers.

Fig 1.8 Convolutional Neural Networks

Recurrent Neural Networks

Recurrent neural networks are used when predictions based on sequential data are
required. Sequential data can be a series of words, images, or other elements. With
the exception of the fact that the layers also get a time-delayed input from the prior
instance prediction, the RNN's

9|Page
topology is identical to that of a feed-forward network. The RNN cell which serves as
a second Input for each prediction, stores this instance prediction.

Fig 1.9 Recurrent Neural Networks

Deep belief networks
A sort of deep learning technique called deep belief networks (DBNs) aims to solve the
issues with conventional neural networks. They accomplish this by exploiting the
network's layers of stochastic latent variables. These feature detectors and hidden units,
also known as these binary latent variables, are stochastic in nature and can potentially
take on any value within a given range.

Restricted Boltzmann Networks or Autoencoders

Networks called Autoencoders encode input data as vectors. They produce a concealed
or compressed version of the original data. The vectors are helpful for reducing the
number of dimensions; they condense the raw data into a smaller number of crucial
dimensions.

RBM is a two-way translator's mathematical counterpart. A forward pass converts

inputs into a string of numbers that encrypts the inputs. This set of numbers is then
converted back into reconstructed inputs during a backward pass. Back prop is a skill
that a skilled net does with

10 | P a g e
great accuracy. Decoders and Autoencoders work together to enable the reconstruction
of input data from its hidden representation.

Fig 1.10 RBM Structure

Types of Learning

Based on the methods and way of learning, machine learning is divided into mainly four
types, which are [3].

i. Supervised Learning
ii. Unsupervised Learning
iii. Semi-Supervised Learning
iv. Reinforcement Learning

Deep Learning
This branch of artificial intelligence is a method that takes its cues from how the human
brain processes data. It relates to applying lessons from instances. A computer model
uses DL systems to help it filter the input data through layers of prediction and
classification. Deep learning

11 | P a g e
utilizes similar information processing techniques to the human brain. Technologies like
driverless automobiles use it. Convolutional neural networks, recurrent neural networks,
and recursive neural networks are the three types of DL network designs.

Supervised Learning
The foundation of supervised machine learning is supervision. In the supervised
learning technique, this means that we train the machines using the "labeled" dataset,
and then the machine predicts the output based on the training. Here, the labeled data
indicates which inputs have already been mapped to which output. To explain it more
precisely, we can say that after training the machine with input and associated output,
we then ask it to predict the output using the test dataset.

The primary objective of the supervised learning method is to correlate the input
variable (x) with the result (y). Risk evaluation, fraud detection, spam filtering, and
other practical uses of supervised learning include these [3].

Classification: Supervised learning problem that involves predicting a class label.

Regression: Supervised learning problem that involves predicting a numerical label.

Fig1.11 Supervised Learning

Unsupervised Learning

12 | P a g e
Because there is no requirement for supervision, unsupervised learning differs from the
supervised learning technique. The system is trained using the unlabeled dataset in
unsupervised machine learning, and it predicts the results without any human
supervision. In unsupervised learning, the models are trained on unlabeled, unclassified
data, and they use that unsupervised data to make decisions.

The unsupervised learning algorithm's primary goal is to classify or group the unsorted
dataset based on commonalities, patterns, and differences. The input dataset is given to
machines with the instruction to uncover hidden patterns.

Clustering unsupervised learning problem that involves finding groups in data.

Density Estimation unsupervised learning problem that involves summarizing the
distribution of data.

Visualization: Unsupervised learning problem that involves creating plots of data.

Projection: Unsupervised learning problem that involves creating lower-dimensional
representations of data.

Fig1.12 Unsupervised Learning

Semi-Supervised Learning
A machine learning algorithm that falls between supervised and unsupervised machine
learning is known as semi-supervised learning. It uses both labeled and unlabeled data

13 | P a g e
sets during the training phase to represent the middle ground between supervised (with
labeled training data) and unsupervised (without labeled training data) learning
methods.

The idea of semi-supervised learning is presented to address the shortcomings of

supervised learning and unsupervised learning algorithms. Using all of the available

data effectively, as opposed to only labeled data as in supervised learning, is the

primary goal of semi-supervised learning.

Fig1.13 Semi- supervised Learning

Reinforcement Learning
A software component that uses reinforcement learning autonomously explores its
surroundings by hitting and trailing, acting, learning from experiences, and increasing
performance. Reinforcement learning operates on a feedback-based method. The
objective of a reinforcement- learning agent is to maximize rewards since the agent is
rewarded for each good activity and penalized for each negative one. Due to the way it
functions, reinforcement learning is used in a variety of disciplines, including multi-
agent systems, game theory, operation research, and information theory. Markov
decision processes provide for the formalization of a reinforcement learning issue
(MDP).

14 | P a g e
 Fig1.14 Reinforcement Learning

Different Types of CNN Models

i. LeNet
ii. AlexNet
iii. ResNet
iv. GoogleNet
v. MobileNet
vi. VGG

Features of CNN
i. In the input image, graphic components like lines, gradients, circles, or even
eyes and faces are extremely successfully recognized by CNN.

ii. Convolutional neural networks are extremely effective for computer vision
because of this feature.

iii. CNN does not require any preparation and can be used straight on an
underexposed image.

15 | P a g e
iv. Convolutional neural networks are feed-forward neural networks with up to 20
nodes.
v. The convolutional layer, a specific kind of layer, is what gives
convolutional neural networks their strength.

vi. Each of the numerous convolutional layers that make up CNN can recognise
more complex shapes when they are stacked on top of one another.

vii. It is feasible to recognise handwritten numbers with three or four convolutional

layers, while it is possible to distinguish human faces with 25 layers.

viii. The goal in this field is to enable machines to perceive the world similarly to
humans and to use this understanding for a variety of tasks, including image and video
recognition, image inspection and classification, media recreation, recommendation
systems, natural language processing, etc [9].

Working of CNN
i. The image is appropriately scaled and supplied into the convolutional layer as
input.
ii. Depending on its size, a filter, neuron, or kernel is placed over some of the
pixels in the input image.

Fig 1.15 Working of CNN

iii. The kernel really moves over the input image, multiplying the filter values by
the image's original pixel values (aka computing element-wise multiplications).

16 | P a g e
iv. Pooling is used, which typically results in a reduction in image dimensions
but not in image channels.

v. Finally, it is subjected to an activation function typically referred to as ReLu

[8].

Steps for the plant leaf disease detection using CNN

i. Image pre-processing and labeling
ii. Neural Network Training
iii. Convolution layer
iv. ReLu Layer
v. Pooling Layer
vi. Flattening
vii. Fully Connected Layer

Image pre-processing and labeling

Pre-processing images commonly involves removing low-frequency background noise,
normalizing the intensity of the individual particle’s images, removing reflections, and
masking portions of images.

Fig 1.16 Bacterial Disease of leaf

17 | P a g e
Fig 1.17 steps of leaf disease detection
Image preprocessing is the technique of enhancing data. Furthermore, procedure of
image preprocessing involves cropping of all the images manually, making the square
around the leaves, in order to highlight the region of interest (plant leaves).

During the phase of collecting the images for the dataset, images with smaller resolution
and dimension less than 500 pixels were not considered as valid images for the dataset.
In addition, only the images where the region of interest was in higher resolution were
marked as eligible candidates for the dataset.

Artificial Neural Network Training

Training the deep convolutional neural network for making an image classification
model from a dataset is proposed. Tensor Flow is an open source software library for
numerical computation using data flow graphs. Nodes in the graph represent
mathematical operations, while the graph edges represent the multidimensional data
arrays (tensors) communicated between them. The flexible architecture allows you to
deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile
device with a single API. Tensor Flow was originally developed by researchers and
engineers working on the Google Brain Team within Google's Machine

Intelligence research organization for the purposes of conducting machine learning and
deep neural networks research, but the system is general enough to be applicable in a
wide variety of other domains as well.

Convolution layer

18 | P a g e
CNN is the first layer to extract features from the input image and it learns the
relationship between features using kernel or filters with input images.

ReLu Layer
ReLu stands for the Rectified Linear Unit for a non-linear operation. The output is ƒ(x)
= max (0, x). We use this to introduce nonlinearity to CNN.

Pooling Layer
It is used to reduce the number of parameters by down sampling and retain only the
valuable information to process further. There are types of Pooling: Max Pooling.
Average and sum pooling.

Flattening
We flatten our entire matrix into a vector like a vertical one. So that it will be passed to
the input layer.

Fully Connected Layer

We pass our flatten vector into input Layer .we combined these features to create
a model. Finally, we have an activation function such as Softmax or sigmoid to classify
the outputs [6].

Use of CNN
i. CNNs are a useful tool for classifying images.
ii. There are many advantages to employing this model as the cutting-edge neural
network. because it may be applied to many other fields and perform important
functions like facial recognition, document analysis, knowledge of the climate, picture
recognition, item identification, etc.

iii. The science sectors have benefited greatly from deep learning, and CNN is the
most well- known application since it achieves the benefits of delivering the highest
levels of performance and efficiency.

iv. A feature is an identifiable and practical finding from the input data that
helps with the required picture analysis.

v. CNN learns the features from the input photos. They typically repeatedly
show up in the data to achieve prominence.

vi. Give the convolution layer's input picture.

vii. Think about convolution using the filters and kernels previously discussed.

19 | P a g e
viii. In order to minimize the dimensions, add a pooling layer.
ix. Add each of these layers several times.
x. Feed the output into a fully linked layer after flattening it.
xi. Now use logistic regression to train the model using back propagation [9].

Advantages of CNN
i. Without any human oversight, image recognition problems with very high
accuracy automatically identify the key aspects.

ii. Sharing of weight.

iii. The primary application of convolutional neural networks is in the processing
of images.
iv. Before CNN, object recognition in photos required manual labor.
v. CNN used the concepts of linear algebra to find patterns in images, which
helped scale the process.

vi. The primary benefit of CNNs over conventional neural networks is that
they automatically identify significant features without human oversight.

vii. Completely convolutional CNNs (FCNs) lack fully linked layers since all of
their learnable layers are convolutional.

viii. It is a simple way to learn how to put together the much activation that would
result in an accurate classification of the provided image.

ix. The cost of computing is frequently very high.

Disadvantages of CNN
i. The location and orientation of objects are not encoded by CNN.
ii. inability to input data geographically invariant
iii. A large amount of training data is needed.
iv. Image Classification Using Different Positions
v. negative examples
vi. Dimensional Frame
vii. additional small drawbacks, such as performance
viii. Due to an operation like maxpool, a convolutional neural network is much
slower.
ix. If the CNN contains multiple layers, training will take a long time if the
machine doesn't ….

20 | P a g e
REFERENCES

- Write references in APA format.

21 | P a g e