PLANT LEAF DISEASE DETECTION USING IMAGE

SEGMENTATION AND MACHINE LEARNING

TECHNIQUES

Project report submitted in partial fulfillment of the requirement for

the degree of Bachelor of Technology

in

Computer Science and Engineering/Information Technology

By

Nitika Sharma (141258)

Shaswat Sood(141260)

Under the supervision of

Dr.Pradeep Kumar Singh

to

Department of Computer Science & Engineering and Information

Technology
Jaypee University of Information Technology Waknaghat, Solan-
173234, Himachal Pradesh
 Certificate

Candidate’s Declaration

I hereby declare that the work presented in this report entitled “ Plant leaf disease detection
using image segmentation and machine learning techniques” in partial fulfillment of the
requirements for the award of the degree of Bachelor of Technology in Computer Science
and Engineering/Information Technology submitted in the department of Computer
Science & Engineering and Information Technology, Jaypee University of Information
Technology Waknaghat is an authentic record of my own work carried out over a period from
August 2015 to December 2015 under the supervision of (Supervisor name) (Designation
and Department name).
The matter embodied in the report has not been submitted for the award of any other degree
or diploma.

Nitika Sharma, 141258

Shaswat Sood,141260

This is to certify that the above statement made by the candidate is true to the best of my
knowledge.

(Supervisor Signature)

Dr. Pradeep Kumar Singh

Assistant Professor (Senior Grade)
Computer Science & Engineering and Information Technology
Dated:

i
 ACKNOWLEDGEMENT

On the submission of our thesis report on “Plant leaf disease detection using image
segmentation and machine learning techniques”, we would like to extend our gratitude and
sincere thanks to our supervisor Dr. Pradeep Kumar Singh, Assistant Professor (Senior
Grade), Department of Computer Science & Engineering and Information Technology
for his constant motivation and support during the course of our work . We truly appreciate
and value his esteemed guidance and encouragement from the beginning. We are indebted to
him for having helped us shape the problem and providing insights towards the solution. And
for providing a solid background for our studies and research thereafter. He has been a great
source of inspiration to us and we thank him from the bottom of our heart. Above all, we
would like to thank all our friends whose direct and indirect support helped us.

Nitika Sharma,141258

Shaswat Sood,141260

ii
INDEX

1 Introduction…………………………………….…………………1-6
1.1Image Basic Concept…………………………………….………1
1.2Neural Network Overview………………………………………1-3
1.3 Pattern recoginisation…………………………………….…….3-6
2 Literature Review…………………………………….…………..7-22
3.Dataset…………………………………….………………………23-27
4.Data Fitting…………………………………….…………………28-32
4.1DataFitting tool…………………………………….……………28-32
4.1.1 Performance…………………………………….…………….29-30
4.1.2 Training State …………………………………….…………...30
4.1.3 Error Histogram…………………………………….…………30-31
4.1.4 Regression…………………………………….……………….31-32
5.Pattern Reorganization…………………………………….………33-38
5.1Introduction…………………………………….………………...33
5.2Neural Network…………………………………….…………….34
5.3 Dataset…………………………………….……………………..34-35
5.4 Train Data…………………………………….………………….35-38
5.4.1Performance…………………………………….………………36
5.4.2Training data…………………………………….……………...36-37
5.4.3 Error Histogram…………………………………….………….37
5.4.4 Confusion Matrix…………………………………….………...37-38
5.4.5Reciever Operating characteristics……………………………...38
6 Clustering Tool…………………………………….……………….39-44
6.1 Neural clustering…………………………………….……………39-40
6.2 Train Data…………………………………….…………………...40-41
6.2.1 SOM Topology…………………………………….……………41-42
6.2.2 SOM neighbor Connection………………………………….…..42
6.2.3 SOM neighbor Distance…………………………………….…...42-43
6.2.4 SOM input planes…………………………………….………….43
6.2.5 SOM samples hits…………………………………….………….43-44
6.2.6 SOM weight position…………………………………….……… 44
References…………………………………….………………………..45-46
Appendices…………………………………….……………………….47-52

iii
 LIST OF ABBREVIATIONS

DCT Discrete Cosine Transform

DWT Discrete Wavelet Transform
GIF Graphics Interchange Format
TIFF Tagged Image File Format
DES Data Encryption Standard Algorithm
LSB Least Significant Bit Algorithm
RSA Rivest-Shamir-Adleman
WMSW Wireless Multimedia Sensory Networks
DPCM Differential Pulse Code Modulation
CR Compression Ratio
MSE Mean Square Error
PSNR Peak Signal to Noise Ratio
CS Compressed Sensing

iv
LIST OF FIGURES

1.1 Neural Network System

3.1 Matlab Image Output1
3.2 HIS Component
3.3 Matlab Image Of diseased
3.4 HIS of diseased
4.1 Nf tool
4.2 Nntrain tool
4.3 Plot Perform
4.4 Plottrainstate
4.5 Ploterrhist
4.6 plotRegression
5.1 NPR Toolbox
5.2 NN layers
5.3 nntrain tool for pattern
5.4 plotperform for pattern
5.5 plottrainstate for pattern
5.6 Ploterrhist in pattern
5.7 plotconfusion for pattern
5.8 ROC characteristics
6.1 nctool
6.2 Nntrain tool for cluster
6.3 Plotsomtop
6.4 Plotsomnc
6.5 plotsomnd
6.6 Plotsomplanes
6.7 plotsomhits
6.8 plotsompos

v
 LIST OF TABLES

3.1 HIS and Histogram Readings

3.2 Target data set
3.3 Test inputs
5.1 Input for pattern tool
5.2 target for pattern tool

vi
 ABSTRACT

A huge development has been made in the field of image processing and
machine learning and its application in various branches of engineering. We
have entered the era of digitization. We have captured images with the help of
digital camera. More clear the images are better and efficient the results. In this
report we have done the classification of disease free, partially diseased and
completely diseased leaves. We have used HSI color model for classification of
our attributes and further we have used Neural Network Toolbox in Matlab for
machine learning and analyzing the results.

vii
 CHAPTER-1
INTRODUCTION

1.1 Image Basic Concept

Image is a collection of rectangular array of dots which are known as pixels. The size of an
image can be determined by the number of pixels present in it. It can simple be calculated by
width x height. Each and every pixel present in an image is a certain type of color. While
working with the black and white image in which the pixels are totally white or totally black,
the options are really restricted as only a single bit is needed for every pixel. Such type of
images are useful for line art like cartoons in newspaper. An additional type of colorless
image is a grayscale image. Grayscale images are very often incorrectly termed as "black and
white" images. They use 8 bits per pixel, which are sufficient enough to depict every shade of
gray color that a human eye can recognize.

While we are working with the color images, things tend to become slightly complicated. The
number of bits per pixel is referred as the depth of an image or the bit plane of an image. A
bit plane which is having n bits can have 2n colors present in it. The human can identify
around 224 colors and some claim the larger number as well. Very commonly found color
depths are 8,16 and 24.

There are two basic methods to store the information of various colors present in an image.
The most direct way is to use the RGB (red, green, blue) color composition in which color of
each pixel is represented by giving an order triple of numbers.
The second way to store information regarding the color is by using the table in order to store
the information of triples and use a reference in the table for every pixel present. This helps in
the betterment of storage requirements of an image.

1.2 Neural networks overview

Neural network are made out of basic components . These components are handled by the
nervous system of the human body. As in nature, the components are associated in such a
way that they operate the working of the system to large extent. You can prepare a neural
system to play out a specific capacity by changing the estimations of the associations
(weights) between components.

Regularly, neural network are balanced, or prepared, with the goal that a specific info
prompts a particular target yield. The following figure shows such a circumstance. Here, the
system is balanced, in light of an examination of the yield and the objective, until the point
that the system yield coordinates the objective. In order to train a network the input/target are
necessary.

1
 Figure 1.1

Neural network systems are designed in such a way that they solve difficult problems very
easily containing design acknowledgment, ID, grouping, discourse, vision, and control
frameworks.

Neural systems can likewise be prepared to take care of issues that are troublesome for
ordinary PCs or individuals. The tool compartment stresses the utilization of neural system
standards that development to—or are themselves utilized as a part of—building, money
related, and other down to earth applications.

The accompanying points disclose how to utilize graphical instruments for preparing neural
systems to tackle issues in work fitting, design acknowledgment, bunching, and time
arrangement. Utilizing these instruments can give you a great prologue to the utilization of
the neural system toolbox programming:

1.2.1 Procedure of the working of NN toolbox:

Four methods for operating NN toolbox.

Apparatuses are applied by principal route. Functionalities of these tools can be accessed
from an ace device began by the command nnstart.

These apparatuses give a helpful method to get to the capacities of the tool compartment for
the accompanying assignments:
• Function used for fitting is: nftool
• Function for recognizing of pattern: nprtool
• Function for clustering of data: nctool
• Function for analyzing the series of time: ntstool

The consecutive method for the utilization of NN toolbox is with the help of essential order
linear tasks. Summon linear activities provide excessive adaptability as compared
apparatuses. On the off chance that this is your first involvement with the tool compartment,
the instruments give the best presentation. What's more, the devices can produce contents of
recorded matlab code in order to give layouts to make tweaked order linear capacities. The
way toward utilizing the instruments to begin with, and afterward producing and altering
Matlab contents, is a magnificent method to find out the usefulness of NN toolbox.

The next method for utilizing the NN tool kit by customization. Such propelled ability
enables us for making our particular custom NN systems, as yet approaching the full

2
usefulness of the tool compartment. You can make systems with subjective associations,
regardless. You have the capacity to prepare them utilizing existing tool kit preparing
capacities .

The fourth method to utilize the tool kit is through the capacity to change any of the
capacities contained in the tool compartment. Each computational part is composed in matlab
code and is completely open.

All subsequent levels of tool stash utilization traverse the fledgling to the master: basic
apparatuses control the new client through particular applications, and system customization
enables analysts to attempt novel models with insignificant exertion.

1.2.2 Neural Network Design Steps

• Firstly do the collection of the data

• Then a network generation takes place.
• Composition of network takes place.
• Declaration of weights and biases takes place.
• Network is being trained.
• Validation of the network takes place.
• Then the network is ready for application.

1.3 Pattern Recognition

Pattern recognition is a fast growing technique now a days. It is playing a very significant
role in various other techniques as well. It is a process with the help of which a pattern is
recognized by computer or machine. It helps in putting patterns in to various categories with
its reliable and efficient methods. The demand of such technique is increasing at a rapid pace.
They can be deployed in real life applications like agriculture, weather forecast ,automatic
disease detection and speech recognition etc.

Pattern recognition is a part of artificial intelligence which helps us in finding the regularities
or a particular pattern present in the data. It is firmly related to technique called machine
learning. Pattern recognition are basically used in vision of computer. It is an essential part
of artificial intelligence that intends t provide human intelligence to the machines or the
computer. It is having the capability to solving complex problems, efficient classification of
the data and solve various other real world problems as well. It supports many versatile fields
like computer science, mathematics and cognitive science etc.

It includes three basic elements which are features, pattern and classifier. Feature means the
prominent attribute or a characteristic of data of an image. Features can include the numeric
data like height or the color. Classifiers help in the division of the feature space into various
decision regions. A decision region is separated by a decision boundary. A general pattern
recognition system is composed of a sensor, a mechanism for preprocessing , a mechanism
for feature extraction, classification algorithm and a training set.

Sensor:
It is an equipment which is used to recognize the actual physical object. If gives output
usually in digital form so that in can be easily processed by the machine. The sensor is
usually chosen from the sensors which are already existing.

3
Pre-processing:

It helps in the production of an efficient set of data by doing noise filtering, smoothing and
normalization. It usually processes larger amount of data and reduces the various variations
present in it. It helps in safe keeping of an image from various errors.

Feature Extraction:

It is basically used to collect the required information from the data input used by the sensor
so that the classification can be done easily. It is usually done with the help of a software
which can be modified according to the sensor.

Classification:

It is a technique which is used to do the classification of the object based upon its properties.
It uses the various features which have been classified by the feature extraction and assign it
to various classes according to its attributes. There are various categories of classification like
nearest mean classification, classification using feed forward artificial neural network etc
which are being used according to the requirement.

1.3.1 Pattern Recognition Technique:

Pattern recognition comprises of the following three things:

• Preprocessing
• Feature Extraction
• Classification

The foremost priority is to get a database which totally depends on the application. After the
acquisition of the appropriate database it is pre-processed so that in can be efficiently used for
the further steps. Features which have been carefully extracted from the database are then
converted to the feature vectors. These feature vectors help in the statistical representation of
the data. So, according to the application domain the classification of these features is done.

1.3.2 Preprocessing:
It is the initial step which is being performed when we do pattern recognition. It is quite
effective for the further stages of the pattern recognition. It helps in improving the
performance and efficiency of the system. It helps in the production of the consistent set of
data by reducing various inconsistencies present in it. It protects the image from various
errors .It helps in the doing the image segmentation i.e. dividing the image into smaller parts
as per requirements. Segmentation plays a very critical role when it comes to the infection
detection in the agricultural products.

1.3.3 Feature extraction

Feature extraction helps in solving the problem of high dimensionality of the input set which
we are providing in pattern recognition. It helps in a decreasing the representation of set of
features as it transforms the data into the feature vector. It only extracts the useful
information from the input which is being provided so that we can get the desired results. The
computation process of the extracted features is quite simple and it doesn't respond to various

4
distortions and variations occurring in the images. Then the features which provide most
exact and favorable outcome are selected. feature extraction can be implemented by various
techniques like Fourier transform, Fuzzy invariant vector, Gabor transform and radon
transform etc.

a) Fourier transform:

Fourier transform is basically used to examine the signal in order to check its frequency. It
has properties like rotational property and translation property. It takes into account the
spectrum magnitude and excludes the circular shift effect.

b) Fuzzy invariant vector:

After the feature has been extracted, it is then converted to a fuzzy invariant vector which
helps in lowering down the effect of the noise which is occurring due to lower frequency. It
also increases the discrimination. The computation of the fuzzy invariant vector is done using
the fuzzy numbers. In this each and every harmonic of a pattern of input has a identical
distribution.

c) Gabor wavelets transform:

This transform helps in the analysis time and frequency parameters. This transformation is
based on the wavelets which is useful for the feature extraction. It also provides effective
resolution. for doing effective pattern recognition it extracts the features of the input data
locally. It works in three domains. These three domains are biological, empirical and
mathematical. Its resemblance to the human vision system provides excellent results.

d) Radon transformation:

In this type of transformation the mapping is done by the coordinate system. Mapping is done
with the help of the Cartesian coordinates and the polar coordinates. It helps in doing the
projection of the image with the help of certain angles. The final result of the projection is the
sum of the various intensities of the pixels in the certain direction. By projecting the image
into different orientation slices, the transform does the capturing of the features quite
effectively. This type of transform can be very well implemented in the Fourier domain.

1.3.4 Classification

The versatile set of required features extracted from various patterns in the past stages are
utilized here. Here the classification and recognition of the features is done and they are
mapped to their respective classes. Learning procedures are categorized into two parts. One is
the supervised learning and other one is unsupervised learning. In the case of supervised
learning, the classifiers are very well aware of the each and every pattern category among
various pattern classes. Where as in the case of the unsupervised learning , the various
attributes of the system are modified on the basis of input given to the system. It searches for
the similar patterns in the data and find the correct output values.

a) Fuzzy ART:
ART stands for Adaptive Resonance Theory. It very well adaptable with the brain of human
beings while doing the processing of the information. It can very well remember the

5
information without forgetting the information which is previously stored. Here the
distribution of the various patterns is done on the basis of the previously stored patterns. A
new pattern is formed , if no resemblance is found from the existing patterns.

b) Neural Networks:
Neural networks are mainly based on the biological concept in order to do the recognition of
the patterns. It is very effective and powerful tool in order to achieve higher performance. It
gathers the information from the human brain. The classification if done by mapping the
feature space with the resultant classes. It is acts as a link between the input as well as the
output sets. For better results multiple set of neural networks can be used. By using various
combining methods input pattern is classified.

c) Markov random field:

MRF is used to combine the statistical data and the information of a particular structure. It
extracts the states which are highly effective and with the help of this they design the
statistical data.

d) Support Vector Machine

The Support Vector Machine classifier can deal with directly divisible information and in
addition non-straightly distinct information utilizing part works. The piece capacity, for
example, polynomial, Gaussian outspread premise work, exponential spiral premise work,
spline , wavelet and autocorrelation wavelet bit, can outline preparing cases in input space
into a highlight space.

e) Multi-class SVM
Multi-class SVM can be favored as a meta-level student. Multi-Classifier framework is
higher order precision, in view of SVM for design acknowledgment. This combinational
technique classifier depends on stacked speculation which consolidate classifiers from
various students, having a two-level structure.

Example acknowledgment strategies are utilized as a part of horticulture applications. A

product model framework for rice sickness discovery is produced by utilizing design
acknowledgment strategies . Its database comprises of the contaminated pictures of different
rice plants which handled by picture developing, picture division methods to distinguish
contaminated parts of the plants. At that point the tainted piece of the leaf has been utilized
for the arrangement reason utilizing neural system.

6
 CHAPTER-2

LITERATURE REVIEW:

2.1 Paper Review: S.Phadiar ,J.Sil, Rice Disease Identification using Pattern Recognition
Techniques, Proceedings of 11th International Conference on Computer and Information
Technology (ICCIT 2008), 25-27 December, pp.420 - 423, 2008.

Objective:
The aim of this paper is to describe a software prototype system for the detection of disease
in rice plant on the basis of various images of the rice plants. Images of the infected part of
the rice plant are taken using digital camera. In order to detect the defected part of the plant
various techniques like image segmentation, image growing etc. have been used. By using
neural network the infected part f the leaf is classified. Image processing and soft computing
techniques have been applied on infected rice plant.

Techniques/Methodology adopted in paper:

• Image processing and pattern analysis methods.

• Hue Intensity Saturation (HIS) model

• Bi-level thresholding method

• Boundary detection algorithm using 8- connectivity method

• Self organizing map (SOM)

Findings/Results:
• In this research paper , the infected part of the rice plant is being classified using
SOM(Self Organizing Map)neural network where the images are being obtained by
doing the extraction of the infected part while four different types of images are being
used for the testing purposes.

• Usage of zooming algorithm is also there for feature extraction of the image. Zooming
algorithm by the usage of computationally efficient technique extracts the features of
the image.

Limitation of the Proposed schemes:

• Results of zooming algorithm are not up to the mark and require some improvement..
Results shown by this are not that much extraordinary infact they are just satisfactory.

• It has also been observed that transformation of image in the frequency domain does
not give a better classification when compared with the original image.

Future Scope:
By using efficient pattern recognition techniques, the system will be able to do the timely
diagnosis of the field problem and the suggestion will help the farmers to take the appropriate
measure to increase the quality of the crop .It will not only reduce the development cost in the
future but also save the environment also.

7
2.2 Paper Review: A. Meunkaewjinda, P. Kumsawat and K. Attakitmongcol, Grape leaf
disease detection from color imagery using hybrid intelligent system, 5th International
Conference on Electrical Engineering/Electronics, Computer, Telecommunications and
Information Technology, Volume: 1,pp. 513 - 516,2008

Objective:
The aim of this paper is to do automatic plant disease diagnosis with the usage of multiple
artificial intelligent techniques. In this paper the main focus is on the grape leaf disease. Once
the system is trained, it can diagnose the plant leaf disease without doing its maintenance
again and again from the beginning.

Techniques/Methodology adopted in paper:

• Genetic algorithm for optimization.
• Support vector machines for classification.
• Artificial neural network.
• Back-propagation neural network(BPNN).
• Anisotropic diffusion technique.
• Modified self-organizing feature map(MSOFM)

Findings/Results:
• Grape leaf disease extraction and classification system using color imagery, the
system gives the desirable results. Back-propagation neural network provides efficient
grape leaf extraction with complex background where as Modified self-organizing
feature map and Genetic algorithm provides automatic adjustment for the colour
extraction of the diseased grape leaf.

• This system has tested images of 426x568 pixels. There were 497 scab disease
samples, 489 rust disease sample and 492 non disease samples used to train the
SVMs. For testing stage there were 39 scab disease images,41 rust disease images and
35 non-disease images. The results show that the system provides desirable
performance.

Limitation of the Proposed schemes:

• There were some problems for doing extraction of ambiguous colour pixels from the
background of the image.
• Neural network do not allow better segmentation of the grape leaf disease pixels.

Future Scope:
The system will demonstrate automatic diagnosis capability with very effective performance
for the further agricultural product analysis/inspection system development.
More computational effort will definitely help in the classification and recognition of the
experimental results.

8
2.3 Paper review: E.Omrani, B.Khoshnevisan, S.Shamshirband, H.Saboohi, N.B.Anuar,
M.H.N.M.Nasir, 'Potential of radial basis function-based support vector regression for apple
disease detection', Department of Biosystem Engineering, pp.2-19,2014.

Objective: The aim of this paper is to classify disease using soft computing approaches,
Artificial Neural Networks(ANNs) , Support Vector Machines(SVM) in apple.

Techniques/Methodology adopted in paper:

• K-means clustering
• Artificial neural networks (ANNs)
• Support vector machines (SVMs)
• Gray level co-occurrence matrix (GLCM)
• Polynomial based (SVR_Poly)
• RBF-based SVR (SVR_rbf)
• Wavelet transform
• Principal component analysis (PCA)
• Back-propagation neural network.

Findings/Results:
• The usage of K-means and ANNs for clustering and classifying diseases affecting the
leaves of plant. The outcome is 94% correct and swifter by 20% .
• K-means cluster divided the images into two groups: infected and healthy leaf areas.
The diseases were classified based on the extracted features.
• The SVR_rbf model had a very small RMSE (0.13) during training and the value was
0.2 in testing.
• The SVR_poly had RMSE of 0.39 in training and RMSE of 0.42 during testing. It
was seen that SVR_rbf model showed consistently good correlation throughout
training and testing.
• A comparison of SVR_rbf results with SVR_poly and ANN reveals that SVR_rbf
outperforms the POLY model in terms of prediction accuracy.

Limitation of the Proposed schemes:

• The Artificial neural networks (ANNs) approach didn't provide accurate results for
the disease classification.
• The results showed by SVR-polt in detecting apple leaf diseases were not exact.

Future Scope:

• Usage of SVR algorithm should be increased as this algorithm contains the quadratic
programming function resolution which is a work function that leads to a unique,
optimum, and comprehensive solution.
• The SVR approaches against the ANN results demonstrated interesting improvement
in the prediction system. It is potentially a promising alternative to existing prediction
models.

9
2.4 Paper Review: A.Singh,B.Ganapathysubramanian, A.K.Singh,S.Sarkar, Machine
Learning for High-Throughput Stress Phenotyping in Plants,Trends in Plant Science, Volume
21, Issue 2, February 2016, Pages 110-124.

Objective:
The aim of this paper is to give us an overview regarding the work done in the field of plant
stress phenotyping using Machine Learning, classification, quantification and prediction. It
will also tell about the general issues in Machine Learning strategy.

Techniques/Methodology adopted in paper:

• High-throughput phenotyping (HTP)
• High-throughput stress phenotyping (HTSP).
• ML algorithms
• Support vector machines (SVM)
• Artificial Neural Networks(ANN)

Findings/Results:
This review gave us an overview of Machine Learning and with the various advantages of
machine learning in the future. The concepts discussed here can be applied to data collected
across the spectrum of complexity and sophistication. We have identified several future
avenues for using ML techniques that show tremendous promise but remain currently
unutilized by the phenotyping community.

Limitation of the Proposed schemes:

• The features identified in the unsupervised process may not be meaningful to a human
user.
• Generative model does not give better results as compared to the discriminative
model.
• Discriminative model is less robust when it comes to over fitting issues.

Future Scope:
• Machine Learning approaches are scalable and also can provide modular strategy for
the data analysis especially for the new domain of 'plant stress analysis'.
• It will also help in the gene discovery process as well as the introduction of novel
selections protocols for the complex competitive traits like biotic and abiotic stress
and yield.

10
2.5 Paper Review: A.Camargo, J.S.Smith, Anximage-processing basedxalgorithm to
automaticallyxidentify plant disease visualxsymptoms, BiosystemsxEngineering ,Volume
102, Issue 1, January 2009, pp. 9-21.

Objective:

The aim of this paper is to do the automatic identification of the plant disease by image
processing from the visual symptoms by analyzing the colored images.

Techniques/Methodology adopted in paper:

• Image pre-processing
• Image enhancement
• Image segmentation
• Image post-processing

Findings/Results:

• The test set consisted of 20 images which were showing symptoms of plant disease
in different crops used in the study. To create the manually segmented set of images, a
grid was overlaid on the image and each position was then evaluated the white colour
and black colour. White colour (1)depicted the pixel having diseased symptoms
whereas the black (0) for non-diseased region.

• To evaluate the algorithm, original images were automatically segmented. The output
which was produced was a binary image where 1 represented a pixel classified as
diseased and 0 as non diseased.

Limitation of the Proposed schemes:

• Huge variation was there in results. To develop such detection system is a very
difficult and challenging task.

• Accuracy is a major criteria in disease detection system.

Future Scope:

• ThexstrengthxofxthisxalgorithmXisxitsxabilityxtoxidentifyxthexcorrectxtargetx(disea
sedxregion) which is shown in the imagesxwithxdifferentxrangexof
intensitiesxdistribution which will definitely help in the future.

11
2.6 Paper review: S.Bashir,N.Sharma,xRemote Area Plant diseasexdetection usingxImage
Processing’, IOSR Journal of Electronicsxand Communication Engineering , Volume 2, Issue
6 ,pp.31-34,2012.

Objective of Paper: Disease can be recognized by using color and texture features. Disease
detection in Malus Domestica is using K mean clustering ,color and texture analysis.

Techniques/Methodology adopted in paper:

• Otsu Segmentation
• K-mean clustering
• Back propagation
• Neural Network
• Co-occurrence matrix method

Findings/Results:
• Appropriate enhancement of images uses histograms. Image segmentation is used for
presence of adequate symptoms for detection of disease.
• Spot on the image can be detected by texture segmentation. Rough, silky, bumpy
texture of image can be identified by texture analysis. Texture analysis uses co-
occurrence matrix method ,which uses Hue Saturation Intensity color space
representation.
• Colorfulness in HIS space is given by the saturation component and transformation of
color space can be done easily.

Limitation of the Proposed schemes:

• It is only limited to symptoms of particular disease not all.
• Large training sets required to recognize various leaves with pest or damaged leaves
due to insects or diseases.

Future Scope:
Conventional method of disease detection in plants using naked eye was cumbersome and not
much effective but by using computer vision toolbox the disease detection in plants is less
time consuming and more efficient .

12
2.7 Paper review:S.Arivazhagan,xR.N.Shebiah, S.Ananthix,S.V.Varthini ,Using texture
detection features, recoginizing unhealthy region of plantxleaves and classificationxof plant
leaf disease, Agric Eng Int: CIGR Journal, Vol. 15, No.1,pp.211-217,2013

Objective of Paper: The aim of the paper is to do detection of unhealthyxregion of plant

leaves andxclassification of plant leaf disease using texturexfeatures.

Technique adopted in paper:

• RGB image acquisition
• Color transformation
• Masking and removal
• Mapping of RGB
• Segment the components
• Obtain the useful segments
• Computation of texture
• Classifier

Findings /Results:
• Green color pixels identified based on specified value of threshold . If green
component is lessxthan the thresholdxvalue ,the redx,green andxblue component of
the pixelxis assigned zero and pixels are completely removed.
• Patch size of 32X32pixels chosen such that significant information is not lost.
• Various formulas for Contrastx,Energy,Localxhomogeneity,Cluster shade and cluster
prominence .
• Minimum distance criteria is used for classification phase, the c-occcurence features
forxthe leavesxare comparedxwith thexcorrespondingxvalues inxfeature library.
• Finite set of elements was drawn by using Support Vector Machine(SVM).
• For training the system 5% of the leaf images from each group are used and
remaining serves as the set for testing.

Limitation of the proposed schemes:

• Large number of training samples are needed for disease identification rate.
• Image of the diseased leaf should be properly captured.

Future Scope:

Little computational effort can be used to classify and recognize the experiment results.
Training sample can be increased to improve disease identification rate .

13
2.8 Paper Review: J.Behmann ,A.K.Mahlein ,T.Rumpf , C. Romer ,L.Plumer ,A review of
advanced machine learning methods for the detection of biotic stress in precision crop
protection, Springer Media New York, Precision Agriculture,Volume 16, Issue 3, pp 239–
260,June 2015

Objective of Paper: Biotic stress detection in precision crop protection using advanced
machine learning methods.

Techniques/Methodlogy adopted in paper:

• Support vector machine

• Support vector regression
• Neural networks
• HSV-color space,(Hue ,Saturation,Value)
• Image Segmentation
• K-Means clustering
• RGB Color

Findings /Results:

• Kernel functions used to map the data requiring non-linear discrimination functions
and in this feature space the data is linearly separable. It is used by support vector
machine.
• Non-Linear SVM and Neural Networks gives good prediction accuracies than linear
approaches. On the other hand, best classification performance in the study was given
by the SVM compared to Neural Network.
• Grey scale image is used for extraction of texture features.
• High dimensional data is analysed with unknown statistical characteristics for
precision crop protection by using machine learning methods.

Limitation of the proposed schemes:

• Stress effect of weeds and nitrogen in maize Neural Network gives accuracy upto
(69%to 58%)
• Non-Linear Support Vector Machine gives accuracy upto 85% and is superiror to a
linear SVM(Support Vector Machine)
• Algorithms used are very specific and generalizations of the prediction accuracies are
not justified.

Future Scope:

Non-relevant or redundant features lead to decrease significantly the prediction accuracy and
reducing these number of features is an important step in data analysis.

14
2.9 Paper Review:J.G.A.Barbedo,Digital image processing techniques for detecting,
quantifying and classifying plant diseases, pp.1-12,Barbedo SpringerPlus ,2013

Objective of Paper: Detecting, quantifying and classifying plant diseases using digital image
processing techniques.

Techniques/Methodology adopted in paper:

• Neural Networks
• Thresholding
• Dual-segmented regression analysis
• Quantification
• Colour analysis
• Fuzzy Logic
• Knowledge-based system
• Sobel operator
• Chlorosis algorithm

Findings/ Results:

• Background is discriminated from the leaf and then damaged regions is separated
from healthy surface and the ratio between the number of pixel in damage by the total
number of pixel of the leaf gives the final estimate.
• Subsequent steps uses 2 modification versions of I3 and only one modification of H.
Sepration of diseased and healthy region is done using thresholding.
• Red and green component of image are combined using chlorosis algorithm for
determining the yellowness of leaf. To discriminate leaves from background blue
component is used.To identify and quantify the necrotic region is done Necrosis
algorithm.
• Area occupied by the spots is estimated using thresholding the blue component of the
image and algorithm to implement using white spots algorithm.

Limitation of the proposed schemes:

• Real-time monitoring was not used.

• Over fitting, overtraining, undersized sample sets and sample sets with low
representativeness are the major problems.

Future Scope:

Crops are continuously monitored by the Real-time monitoring and alarm willbe issued as
soon as disease is detected.

15
2.10 F.Qin, D.Liu, B.Sun, L.Ruan, Z.Ma, H.Wang,Identification of Alfalfa Leaf Diseases
Using ImageRecognitionTechnology,p.p.1-15,Plos Journals,December 15, 2016

Objective of Paper:Using Image recognition technology identification of Alfalfa leaf

diseases.

Techniques/Methodology adopted in paper:

• Fuzzy C-MEANS CLUSTERING

• K-MEDIAN CLUSTERING
• Euclidean DISTANCE
• Logistic REGRESSION ANALYSIS
• Naive BAYES ALGORITHM
• Cart
• LINEAR DISCRIMINANT ANALYSIS

Findings/ Results:

• Arithmetic square root of total number of features was randomly selected by each
decision tree .For eg. If arithmetic square root is decimal ,then rounding up the
decimal gives the number of features randomly selected by each decision tree.
• Disease recognition models built after feature selection gives the satisfactory
recognition results. This indicates that features extracted from lesion images were
efficient.
• Relief method gives the top 45 features for importance ranking based on the SVM
model.
• For implementing K-median clustering algorithm linear discriminant was used and the
highest score of median and mean are used for implementation.

Limitation of the proposed schemes:

• To get accurate results there is need of large datasets.

• Accuracy and efficiency depends on the experience and it is time consuming and
subjective work.

Future Scope:

Optimal image recognition model of alfalfa leaf diseases can be done by developing mobile
application.

16
2.11 Paper Review: A.Meunkaewjinda, P.Kumsawatxand K.Attakitmongcol,Grapexleaf
diseasexdetection from color imageryxusing hybrid intelligentxsystem, 5th International
Conference on Electrical Engineering/Electronics, Computer, Telecommunications and
Information Technology, Volume: 1,pp. 513 - 516,2008

Objective: The aim of this paper is to do automatic plant disease diagnosis with the usage of
multiple artificial intelligent techniques. In this paper the main focusis on the grape leaf
disease. Once the system is trained, it can diagnose the plant leaf disease without doing its
maintenance again and again from the beginning.

Techniques/Methodology adopted in paper:

• Genetic algorithm for optimization.
• Support vector machines for classification.
• Artificial neural network.
• Back-propagation neural network(BPNN).
• Anisotropic diffusion technique.
• Modified self-organizing feature map(MSOFM)

Findings/Results:
• Grape leaf disease extraction and classification system using color imagery, the
system gives the desirable results. Back-propagation neural network provides efficient
grape leaf extraction with complex background where as Modified self-organizing
feature map and Genetic algorithm provides automatic adjustment for the colour
extraction of the diseased grape leaf.

• This system has tested images of 426x568 pixels. There were 497 scab disease
samples, 489 rust disease sample and 492 non disease samples used to train the
SVMs. For testing stage there were 39 scab disease images,41 rust disease images and
35 non-disease images. The results show that the system provides desirable
performance.

Limitation of the Proposed schemes:

• There were some problems for doing extraction of ambiguous colour pixels from the
background of the image.
• Neural network do not allow better segmentation of the grape leaf disease pixels.

Future Scope:
• The system will demonstrate automatic diagnosis capability with very effective
performance for the further agricultural product analysis/inspection system
development.
• More computational effort will definitely help in the classification and recognition of
the experimental results.

17
2.12 Paper review: E.Omrani, B.Khoshnevisan, S.Shamshirband, H.Saboohi, N.B.Anuar,
M.H.N.M.Nasir, 'Potential of radial basis function-based support vector regression for apple
disease detection', Department of Biosystem Engineering, pp.2-19,2014.

Objective: The aim of this paper is to classify disease using soft computing approaches,
Artificial Neural Networks(ANNs) , Support Vector Machines(SVM) in apple.

Techniques/Methodology adopted in paper:

• K-means clustering
• Artificial neural networks (ANNs)
• Support vector machines (SVMs)
• Gray level co-occurrence matrix (GLCM)
• Polynomial based (SVR_Poly)
• RBF-based SVR (SVR_rbf)
• Wavelet transform
• Principal component analysis (PCA)
• Back-propagation neural network.

Findings/Results:

• The usage of K-means and ANNs for clustering and classifying diseases affecting the
leaves of plant. The experimental results revealed that proposed approach can
successfully detect and classify the inspects disease with 94% precision and 20%
faster.

• K-means cluster divided the images into two groups: infected and healthy leaf areas.
The diseases were classified based on the extracted features.

• The SVR_rbf model had a very small RMSE (0.13) during training and the value was
0.2 in testing.

• The SVR_poly had RMSE of 0.39 in training and RMSE of 0.42 during testing. It
was seen that SVR_rbf model showed consistently good correlation throughout
training and testing.

• A comparison of SVR_rbf results with SVR_poly and ANN reveals that SVR_rbf
outperforms the POLY model in terms of prediction accuracy.

Limitation of the Proposed schemes:

• The Artificial neural networks (ANNs) approach didn't provide accurate results for
the disease classification.
• The results showed by SVR-polt in detecting apple leaf diseases were not exact.

Future Scope:

18
• Usage of SVR algorithm should be increased as this algorithm contains the quadratic
programming function resolution which is a work function that leads to a unique,
optimum, and comprehensive solution.

• The SVR approaches against the ANN results demonstrated interesting improvement
in the prediction system. It is potentially a promising alternative to existing prediction
models.

19
2.13 Paper Review:A.Singh,B.Ganapathysubramanian, A.K.Singh,S.Sarkar, Machine
Learning for High-Throughput Stress Phenotyping in Plants,Trends in Plant Science, Volume
21, Issue 2, February 2016, Pages 110-124.

Objective:

The aim of this paper is to give us an overview regarding the work done in the field of plant
stress phenotyping using Machine Learning, classification, quantification and prediction. It
will also tell about the general issues in Machine Learning strategy.

Techniques/Methodology adopted in paper:

• High-throughput phenotyping (HTP)

• High-throughput stress phenotyping (HTSP).
• ML algorithms
• Support vector machines (SVM)
• Artificial Neural Networks(ANN)

Findings/Results:

This review gave us an overview of Machine Learning and with the various advantages of
machine learning in the future. The concepts discussed here can be applied to data collected
across the spectrum of complexity and sophistication. We have identified several future
avenues for using ML techniques that show tremendous promise but remain currently
unutilized by the phenotyping community.

Limitation of the Proposed schemes:

• The features identified in the unsupervised process may not be meaningful to a human
user.
• Generative model does not give better results as compared to the discriminative
model.
• Discriminative model is less robust when it comes to over fitting issues.

Future Scope:

• Machine Learning approaches are scalable and also can provide modular strategy for
the data analysis especially for the new domain of 'plant stress analysis'.

• It will also help in the gene discovery process as well as the introduction of novel
selections protocols for the complex competitive traits like biotic and abiotic stress
and yield.

20
2.14 Paper Review:VijaiSingh, A.K.Misra, Detectionxofxplant leafxdiseases usingximage
segmentationxand softxcomputing techniques,Volume 4,Issue 1, March 2017, PP. 41-49.

Objective:

The aim of this paper we do the automatic detectionxand classification ofxplant leaf diseases
using an algorithm for imagexsegmentation technique. It also covers surveyxon different
diseases classificationxtechniques that can be used for plantxleaf diseasexdetection.

Techniques adopted in paper:

• Genetic algorithm
• k-nearest-neighbor method
• Machine learning based recognition
• Color Co-occurrence Method
• Artificial neural network (ANN)

Findings/Results:
• By using Minimum Distance Criterion with K-Mean Clustering we did the first
classification which showed itsxefficiency and accuracyxof 86.54% . Thexdetection
accuracy was later improvedxto 93.63% by the proposedxalgorithm.
• The secondxphase ofxclassification was done using SVMxclassifier andxshows
efficiency and accuracyxofx95.71%.But with the help of proposed algorithm we were
able to improve the detection accuracy to 95.71%.
• From thexresults it was clearly visible that onlyxfew sample from the Frogxeye leaf
spot and bacterialxleaf spot leaves werexmisclassified. Only two leafs with bacterial
leaf spot disease were classified as frog eye leaf spot and one frogxeye leafxspot was
classified as the bacterialxleaf spot. The average accuracy of proposed algorithm is
97.6% .

Limitation of the Proposed schemes:

• If the trainingxdata is not linearlyxseparablexthen it becomes quite difficultxto
determine the optimalxparameters in the Support Vector Machine, which
appears to be its major drawback.
• There is a need to do improvement in the recognition rate of the classification
process.

Future Scope:
• Genetic algorithm optimizes continuous and discrete variables effectively. It searches
for large data samples of the cost surface with large variables being processed at the
same time.
• It helps in the optimization of variables with highly complex cost surfaces.

• Neural Networks canxbexused for the recognitionxrate of the classificationxprocess.

21
2.15 Paper Review: A.Camargo, J.S.Smith,xAnximage-processingxbasedxalgorithm to
automaticallyxidentifyxplantxdiseasexvisualxsymptoms,xBiosystemsxEngineering ,Volume
102, Issue 1, January 2009, pp. 9-21.

Objective:

The aim of this paper is to do the automatic identification of the plant disease by image
processing from the visual symptoms by analysing the coloured images.

Techniques/Methodology adopted in paper:

• Image pre-processing
• Image enhancement
• Image segmentation
• Image post-processing

Findings/Results:

• The test set consisted of 20 images which were showing symptoms of plant disease
in different crops used in the study. To create the manually segmented set of images, a
grid was overlaid on the image and each position was then evaluated the white colour
and black colour. White colour (1)depicted the pixel having diseased symptoms
whereas the black (0) for non-diseased region.

• To evaluate the algorithm, original images were automatically segmented. The output
which was produced was a binary image where 1 represented a pixel classified as
diseased and 0 as non diseased.

Limitation of the Proposed schemes:

• Huge variation was there in results. To develop such detection system is a very
difficult and challenging task.

• Accuracy is a major criteria in disease detection system.

Future Scope:
• Thexstrengthxof thisxalgorithmxis its abilityxtoxidentify thexcorrectxtargetx(diseased
xregion) which is shown inxthe imagesxwithxdifferentxrangexofxintensities
xdistribution which will definitely help in the future.
• Due to the higher complexityxof the imagesxused in thisxstudy, the
strategyxproposed here will bexsuitable forxother type ofximages as well whose
xtargets arexdifferent to that ofximages which are showingxdiseasedxplants.

22
 CHAPTER-3
DATA SET
A dataset is an accumulation of related, discrete things of related information that might be
gotten to exclusively or in blend or oversaw all in all element.

An informational collection is sorted out into some kind of information structure. In a

database, for instance, an informational collection may contain a gathering of business
information (names, pay rates, contact data, deals figures, et cetera). The database itself can
be viewed as an informational collection, as can assemblages of information inside it
identified with a specific kind of data, for example, deals information for a specific corporate
office.

Data set used in the project is stored in an excel file and further used for getting results in
Matlab software.
Our data set is a collection of four attributes which are as follows:

• The first component of our dataset comprises of Histogram of the leaf which shows
the characteristics of either of diseased leaf or disease free leaf which further is
analyzed and outcome is generated accordingly.
• The second component consists of the Hue component which is generally shows the
dominant color.
• The third component is S component it shows the purity of the image or addition of
white light.
• The last component is the I-Component it states the amplitude of light present in our
image.

Figure3.1 shows the normal image ,grayscale image, Histogram Equal and the
histogram.

23
 Figure3.2 shows the normal image the H,S and I (Hue,Saturation,Intensity)component
of the good leaf.

Figure3.3 The Histogram of the grayscale ,Histogram Equal image, Histogram, Diseased
leaf.

Figure3.4 It shows the normal image, H-component, S-component and I-Component

of the diseased leaf.

24
Table3.1 This is our data set which consists of Histogram, H-Component ,S-Component and
I-Component respectively. It is fed as input to the neural network.

Our dataset is not redundant and every time whenever the input is fed to the network a unique
result is generated depending upon above stated characteristics.

25
Table3.2 This figure depicts the target of previously given input in order to get the desired
result whether the leaf is prone to disease or not.

Our result is divided into 3 categories that are 1,2 and 3 .They are explained as follows:

• If the value of target is 1 and less than 1.5 then it means that is is a Good leaf which is
free from disease.
• Secondly if the value is 2 and less than 2.5 than it semi diseased leaf in which some
portion is exposed to some kind of infection.
• Lastly if the target value if 3 and greater than 3.5 than it is completely infected.

26
 Table3.3 These are some of the test inputs which are used for analyzing our NNtool
The first column if the histogram followed by H,S, I-Component followed by the Resultant
data.

27
 CHAPTER-4
Data Fitting

Neural systems are great at fitting capacities. Actually, there is evidence that a genuinely
basic NN system can adjust itself in viable capacity.
Fit information utilizing bends, surfaces, and nonparametric techniques

Information fitting is the way toward fitting models to information and breaking down the
exactness of the fit. Specialists and researchers utilize information fitting procedures,
including numerical conditions and nonparametric techniques, to display gained information.

MATLAB gives you a chance to import and envision your information, and perform
fundamental fitting systems, for example, polynomial and spline interjection. You can
perform information fitting intelligently utilizing the MATLAB Basic Fitting device, or
automatically utilizing MATLAB capacities for fitting.

4.1 Data Fitting Tool

Figure4.1 nftool
Analysis: The info vectors and target vectors will be arbitrarily separated into three sets as
takes after:

• 70% of leaf dataset will be utilized for preparing.

28
 • 15% of leaf dataset will be utilized to approve that the system is summing up and to
quit preparing before overfitting.

• The last 15% of leaf will be utilized as a totally autonomous trial of system
speculation.

The standard framework that is used for work fitting is a two-layer feedforward sort out, with
a sigmoid move work in the hid layer and a straight capacity work in the yield layer. The
default number of covered neurons is set to 10.

Figure4.2 nntraintool
Analysis: In this number of Epoch are having 18 iteration, Performance of the diseased and
diseased free leaf dataset is equivalent to 0.0163,Gradient is equivalent to 0.00559 and the
number of validation checks are equal to 6.

4.1.1Performance:

Figure4.3 plotperform

29
Analysis: After analysis of the data set consisting of diseased, semi-diseased and completely
diseased leaves we are getting result at 12 epochs.
plotperform(TR) are basically used to plot the error vs. epoch. It is useful for doing training,
validating and testing the performances exhibitions in preparation storage TR evaluated by
function train.
By and large, the error diminishes after more epochs of preparing, yet may begin to
increment on the approval informational collection as the system begins overfitting the
preparation information.

Performance Plot demonstrates you mean square error flow for all your datasets in
logarithmic scale. Preparing MSE is continually diminishing, so its approval and test MSE
you ought to be occupied with. Your plot demonstrates an impeccable training.
Mean Square Error (MSE) is the mean (average) extent of the squares of the error: i.e., the
separation between the model's approx of your test esteems and the genuine test esteem.
(squaring just changes over things to a flat out esteem as opposed to fiddling with under or
overshooting).

4.1.2Training State:

Figure4.4 plottrainstate
Analysis: After training the data set of leaves we get that Matlab naturally quits preparing
after 6 flops consecutively.
the figure shows the Gradient =0.055899, at epoch 18 and Mu=0.0001,at epoch 18 and
validation checks =6, at epoch at 18.

Training State demonstrates to you some other preparing insights. Gradient is an estimation
of back propagation slope on every cycle in logarithmic scale. Validation falls flat are
emphases when approval MSE expanded its esteem. A considerable measure of comes up
short means overtraining.

4.1.3 Error Histogram:

30
Figure4.5 ploterrhist
Analysis: You can see that while most bungles fall between - 0.4197 and 0.2298, there is a
readiness point with a slip-up of 8 and endorsement centers with goofs of 09 and 11. These
special cases are furthermore evident on the testing relapse plot. The important contrasts to
the point and a goal of 30 and yield near 15.
The blue bars address getting ready data, the green bars address endorsement data, and the
red bars address testing data. The histogram can give you an indication of inconsistencies,
which are data centers where the fit is by and large more repulsive than the bigger piece of
data. It is a brilliant idea to check the irregularities to choose whether the data is terrible, or if
those data centers are not the same as whatever is left of the educational accumulation. If the
irregularities are considerable data concentrates, yet are not in the least like whatever is left of
the data, by then the framework is extrapolating for these core interests. You should assemble
more data that looks like the inconsistency centers, and retrain the framework.
As the quantity of the error occurring in the dataset results is very less hence it shows that the
classification of diseased ,partially diseased and completely diseased leaves have been done
right.

4.1.4 Regression:

Figure4.6 plotregression

31
Analysis: The going with relapse plots show the framework yields concerning centers for
planning, endorsement, and test sets. For a flawless fit, the data should fall along a 45 degree
line, where the framework yields are comparable to the destinations. For this issue, the fit is
sensibly helpful for every single educational accumulation, with R regards for every
circumstance of 0.97 or above. If fundamentally more correct results were required, you
could retrain the framework by clicking Retrain in nftool. This will change the hidden
weights and slants of the framework, and may make an improved framework in the wake of
retraining. Distinctive options are given on the going with sheet.

32
 CHAPTER-5
PATTERN RECOGNITION USING NEURAL NETWORKS

5.1 Introduction

In pattern recognition the inputs are related with various classes. Neural systems are great at
design acknowledgment issues. A neural system with enough components (called neurons)
can order any information with discretionary precision. They are especially appropriate for
complex choice limit issues over numerous variables. The system will be composed by
utilizing the credits of neighborhoods to prepare the system to deliver the right target classes.
Neural systems have demonstrated themselves as capable classifiers and are especially
appropriate for tending to non-linear issues. They are likewise great at perceiving patterns.

Figure5.1 nprtool
Analysis: Here the validation and test data takes place. Here total 120 samples have been
taken into consideration in which 70% of the samples (84 samples) are used for training ,18
samples out of 120 which is 15% of the total sample are used for validation of the sample
whereas rest 15% that is 18 samples are been used for testing of the data. Data is trained till
the point it doesn't give the proper result during the validation and testing is an independent
process.

5.2 Neural Network

Figure5.2 NN layers

33
Analysis: It consists of Input which are 4 in number and the 10 hidden layers and the output
layer which are 3 in numbers
Feed forward systems comprise of a progression of layers. The primary layer has an
association from the system input. Each resulting layer has an association from the past layer.
The last layer creates the system's yield.
Feed forward systems can be utilized for any sort of contribution to yield mapping. A feed
forward connect with one hidden layer and enough neurons in the hidden layers, can fit any
limited information yield mapping issue.
Specific variants of the feed forward organize incorporate fitting (fitnet) and pattern
recognition (patternnet) systems. A minor departure from the feed forward organize is the
cascade forward system (cascadeforwardnet) which has extra associations from the
contribution to each layer, and from each layer to every after layer.
feedforwardnet(hiddenSizes,trainFcn) takes these contentions,
hiddenSizes -Row vector of at least one concealed layer sizes (default = 10)
trainFcn -Preparing capacity (default = 'trainlm')
also, restores a feedforward neural system

5.3 DataSet:

5.3.1 Input Dataset

The input data set consists of 4 x120 matrix representing the static data:

120 samples of 4 elements.

Table5.1 Input for recognition of the pattern.

5.3.2 Target Dataset

The target data set consists of 3 x 120 matrix, representing the static data: 120 samples of 3
elements.

34
Table5.2 Target for pattern tool.

5.4 Train Data

Neural network training tool it consist of Epoch,time,Performance,Gradient,Valiation

Checks. From this we can plot performance, Training State, Error Histogram, Confusion
matrix and receiver operating characteristics.

Figure5.3 nntrain tool for pattern

35
Analysis: In it epoch has 25 iterations ,performance is 0.00576 and the gradient is 0.0118 and
the number of validation checks are 6.

5.4.1 Performance:

Figure5.4 plotperform for pattern.

Analysis: Performance Plot demonstrates you mean square error flow for all your datasets in
logarithmic scale. Preparing MSE is continually diminishing, so its approval and test MSE
you ought to be occupied with. Your plot demonstrates an impeccable training.
Mean Square Error (MSE) is the mean (average) extent of the squares of the error: i.e., the
separation between the model's approx of your test esteems and the genuine test esteem.
(squaring just changes over things to a flat out esteem as opposed to fiddling with under or
overshooting).It is giving optimal result at 15 epochs. It is efficiently classifying the diseased
and diseased free leaves.

5.4.2 Training state:

36
Figure5.5 plottrainstate for pattern.
Analysis: The figure shows the Gradient =0.18351 at epoch 21and validation checks =6, at
epoch at 21.

5.4.3 Error Histogram:

Figure5.6 ploterrhist in pattern.

Analysis: You can see that while most bungles fall between - 0.1316 and 0.1318.It has zero
error at nearly 0.1.It works with the help of floor and ceiling function if the value is negative
that it shows the lower value and on the other hand if the value is to the positive side it raises
the value. The computation of the diseased and diseased free data has been done very
accurately.

5.4.4 Confusion Matrix:

Figure5.7 plotconfusion in pattern.

37
Analysis: It demonstrates the rate of right and of base orders. The green squares refer to the
correct classification of the results of the diseased, partially diseased leaves and completely
diseased leaves where as on the other hand the red squares refer to the misclassification.
In case of training the confusion matrix the sun of the diagonal elements which are
highlighted in the green color sun up to the value of 98.8% which means that it is showing
correct results of the dataset of the diseased and diseased free leaves to large extent whereas
the value in the red square is mere 1.2% which is a very small error value.
In the case of validation of confusion matrix the sun of the diagonal elements which are
highlighted in the green color comprises of the 88.9% of the total value where as the value
highlighted in the red color is 11.1%.
In the case of testing of the confusion matrix the value highlighted in the green color
comprises of 100% and the values highlighted in the red colored box 0% which is a great sign
and highly effective results.
So, the overall confusion matrix the values which are correctly tested for the diseased ,
partially diseased and completely diseased as it comprises of 97.5% are the values which are
having little bit error are approximately 2.5% which is quite less.

5.4.5 Receiver Operating Characteristics:

Figure5.8 ROC characteristics.

Analysis: The shaded region in every axis represent the ROC bends. The plot is being made
between false positive rate and true positive rate.

38
 CHAPTER-6
CLUSTERING TOOL

Neural systems is phenomenal utilization of grouping information. Gathering information by

similarity. For eg. Performing market division by gathering individuals as indicated by their
purchasing designs. Apportioning information identifies with their information mining.
Grouping qualities with related articulation designs.

Cluster analysis includes applying at least one clustering calculations with the objective of
finding concealed examples or groupings in a dataset. Clustering calculations frame
groupings or groups such that information inside a group have a higher measure of closeness
than information in some other group. The measure of comparability on which the bunches
are displayed can be characterized by Euclidean separation, probabilistic separation, or
another metric.

Cluster analysis is an unsupervised learning technique and a vital undertaking in exploratory

information investigation. Well known bunching calculations include:

• Progressive grouping: assembles a multilevel chain of command of bunches by

making a bunch tree

• k-Means bunching: allotments information into k unmistakable groups in view of

separation to the centroid of a bunch

• Gaussian blend models: models groups as a blend of multivariate typical thickness

segments

• Self-sorting out maps: utilizes neural systems that take in the topology and
appropriation of the data.

6.1 Neural Clustering:

39
Figure6.1 nctool
Analysis: The data is being trained using nctool. The data is being accurately trained in order
to understand the methodology and separation between the input samples proved to the
network.

6.2 Train Data:

40
Figure6.2 nntraintool for clustering.

Analysis: nntraintool opens the neural system preparing GUI.

This capacity can be called to make the preparation GUI obvious before preparing has
happened, in the wake of preparing if the window has been shut, or just to convey the
preparation GUI to the front.
System preparing capacities handle all movement inside the preparation window.
To get to extra helpful plots, identified with the present or last system prepared, amid or
subsequent to preparing, click their individual catches in the preparation window.

6.2.1 SOM Topology:

41
Figure6.3 plotsomtop

Analysis: It is hexagonal. In the abovementioned figure each and every hexagon builds
communication network with neurons. The frame is 10x10 so it leads to 100 neurons in the
system. It consists of four attributes in which the information vector and the information
space is four dimensional.

6.2.2 SOM Neighbor Connection:

Figure6.4 plotsomnc
Analysis: plotsomnc(net) plots a SOM layer indicating neurons as dim blue patches and their
immediate neighbor relations with red lines.
This plot bolsters SOM systems with hextop and gridtop topologies, however not tritop or
randtop.

6.2.3 SOM Neighbor distance:

42
Figure6.5 plotsomnd

Analysis: In this figure the blue colored hexagons builds communication network with
neurons. On the other hand red colored lines act as an association with the adjacent neurons.
The pitch dark colors are related to bigger separations and lighter color speak to the littler
separations.

6.2.4 SOM input Planes:

Figure6.6 plotsomplanes
Analysis: plotsomplanes(net) creates an arrangement of subplots. Each ith subplot
demonstrates the weights from the ith contribution to the layer's neurons, with the most
negative associations appeared as blue, zero associations as dark, and the most grounded
positive associations as red.

6.2.5 SOM Sample Hits:

43
Figure6.7 plotsomhits

Analysis: plotsomhits(net,inputs) plots a SOM layer, with every neuron demonstrating the
quantity of information vectors that it groups. The relative number of vectors for every
neuron is indicated by means of the measure of a shaded fix.

This plot underpins SOM systems with hextop and gridtop topologies, however not tritop or
randtop.

6.2.6 SOM Weight Positions:

Figure6.8 plotsompos

Analysis: plotsompos(net) plots the info vectors as green specks and shows how the SOM
arranges the information space by indicating blue-gray spots for every neuron's weight vector
and interfacing neighboring neurons with red lines. plotsompos(net, inputs) plots the
information close by the weights.
The value of weight1 is till 8 whereas the value mentioned in weight 2 is till 4.5. The used in
the clustering is a four dimensional input as the number of input parameters used are four.

44
 REFERENCES
[1]S.Bashir,N.Sharma,Remote Area Plant disease detection using Image Processing, IOSR
Journal of Electronics and Communication Engineering , Volume 2, Issue 6 ,pp.31-34,2012.

[2]S.Arivazhagan,R.N.Shebiah, S.Ananthi,S.V.Varthini ,Using texture detection features,

recoginizing unhealthy region of plant leaves and classification of plant leaf disease ,
AgricEngInt: CIGR Journal, Vol. 15, No.1,pp.211-217,2013.

[3]J.Behmann,A.K.Mahlein,T.Rumpf,C.Romer,L.Plumer, A review of advanced machine

learning methods for the detection of biotic stress in precision crop protection, Springer
Media New Y ork, Precision Agriculture,Volume 16, Issue 3, pp. 239–260, 2015.

[4]J.G.A.Barbedo,Digital image processing techniques for detecting, quantifying and

classifying plant diseases, ArnalBarbedoSpringerPlus , pp.1- 12,2013.

[5] F.Qin, D.Liu, B.Sun, L.Ruan, Z.Ma, H.Wang,Identification of Alfalfa Leaf Diseases
Using ImageRecognitionTechnology,Plos Journals, pp.1-15, 2016.

[6]A.Meunkaewjinda, P.Kumsawat and K.Attakitmongcol,Grape leaf disease detection from

color imagery using hybrid intelligent system, 5th International Conference on Electrical
Engineering/Electronics,Computer,Telecommunications and Information Technology,
Volume: 1,pp. 513 - 516,2008.

[7]E.Omrani, B.Khoshnevisan, S.Shamshirband, H.Saboohi, N.B.Anuar, M.H.N.M.Nasir,

'Potential of radial basis function-based support vector regression for apple disease detection',
Department of Biosystem Engineering, pp.2-19,2014.

[8]A.Singh,B.G.Subramanian,A.K.Singh,S.Sarkar, Machine Learning for High-Throughput

Stress Phenotyping in Plants,Trends in Plant Science, Volume 21, Issue 2, pp. 110-124, 2016.

[9]V .Singh, segmentation 49,2017. [10]A.Camargo, J.S.Smith, An image-processing

based algorithm to automatically identify plant disease visual symptoms,Biosystems
Engineering ,Volume 102, Issue 1, pp. 9-21, 2009.

[11] ,AnInvestigat -ion Into Machine Learning Regression Techniques for the Leaf Rust
Disease Detection Using Hyperspectral Measurement, IEEE Journal of Selected Topics in
Applied Earth Observations and Remote Sensing 9, 2016.

[12]M.K.Tripathi, Recent Machine Learning Based Approaches for Disease Detection and
Classification of Agricultural Products,ComputingCommunication Control
andautomation(ICCUBEA),2016 International Conference , 2016.

[13]M.Jhuria, A.Kumar, R.Borse, Image processing forSmart Farming : Detection of disease

and food grading, Institute of Electrical and Electronics Engineering, 2013. [14] N.
Bhardwaj, G. Kaur, P.K.Singh, A Systematic Review on Image Enhancement Techniques,
Sensors and Image Processing, Advances in Intelligent Systems and Computing (AISC),

45
Springer, pp. 227-235, 2018.

[15] G. Kaur, N. Bhardwaj, P.K.Singh, An Analytic Review on Image Enhancement

Techniques Based on Soft Computing Approach, Sensors and Image Processing, Advances in
Intelligent Systems and Computing(AISC), Springer, pp. 255-265, 2018.

[16] K. Vasudeva, P.K. Singh,Y. Singh, A Methodical Review on Issues of Medical Image
Management System with Watermarking Approach, Indian Journal of Science and
Technology, Vol 9(32), DOI: 10.17485/ijst/2016/v9i32/100188, ISSN: 0974-5645, 2016.

[17] D. Agarwal, A. Gupta, P. K.Singh, A systematic review on Artificial Bee Colony

Optimization Technique, International Journal of Control Theory and Application, Vol.
9(11), pp. 5487-5500, 2016,

[18] A. Sharma, P.K. Singh, P. Khurana, Analytical Review on Object Segmentation and
Recognition, published in proceedings of 6thInternational Conference - Cloud System and
Big Data Engineering (Confluence), 14th - 15thJanuary, 2016, Noida, India, IEEE, pp. 524 –
530,2016.

[19] R. Bhardwaj, P .K. Singh, Analytical Review on Human Activity Recognition in Video,
published in proceedings of 6thInternational Conference - Cloud System and Big Data
Engineering (Confluence), 14th-15thJanuary, 2016, Noida, India, IEEE, pp. 531 – 536,2016.

46
 APPENDICES

Code for Neural fitting:

47
48
Code for pattern recognition:

49
50
Code For clustering:

51
52