IRJET- Artificial Neural Network and Particle Swarm Optimization in Orange Identification

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 03 | Mar 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 820
Artificial Neural Network and Particle Swarm Optimization in Orange
Identification
Mr. J. Jayakumar M.E1, V.S. Adelin Nittas 2
1Assistant professor, Department of Electronics and Communication Engineering, Annai Vailankanni College Of
Engineering, Azhagappapuram, Kanniyakumari, Tamil Nadu, India.
2Final year PG student, Department of Communication Systems, Annai Vailankanni College Of Engineering,
Azhagappapuram, Kanniyakumari, Tamil Nadu, India.
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract- Advancement in technology hasleadtotheuse of
computer science and it’s associated technologyinagriculture
and food science. Fruit industry contributes a major part in
the growth of a nation. Identifyingfoodsthatlookssimilarisof
major concern. And also, the rise in labour costs, shortage of
efficient labourers and the need to improve the processes at
the production side have all put pressure on producers and
processors to demand for a time-saving, economic, consistent
and non-destructive inspection method. In such scenarios,
automation can reduce the costs while increasing the
production efficiency. Automatic fruit grading and sorting
requires the implementation of computer vision systems.
Hybrid approaches that consists of artificial neural networks
and meta-heuristic algorithmscanprovidesatisfyingsolutions
to these problems. This can provide fast and accurate ways in
classifying fruits. This paper deals with the use of artificial
neural networks and Particle SwarmOptimization(ANN-PSO)
in identifying orange varieties.
Key Words: Artificial Neural Network (ANN),
Metaheuristics, Artificial Neurons, Particle Swarm
Optimization, Nature inspired metaheuristics.
1. INTRODUCTION
Fruits play vital part in a balanced diet. They are said to be
good sources of minerals and vitamins and for their role in
preventing related deficiencies. People those who have the
habit of eating fruits as part of their daily diet are generally
less prone to chronic diseases. Fruitsprovidemany essential
nutrients including fibre, potassium, folate and vitamin C.
The nutrients in fruit are vital for maintaining the health of
one’s body. Oranges are relatively hard to classify when
compared with other fruit varieties. Here, we mainly
consider three types of oranges namely Bam, Pavyandi and
Thomson.
The use of Computer Vision Systems in the field of
agriculture has increased substantially in the recent years
for the substantial information it provides about the nature
and attributes of the produce while reducing costs and
guaranteeing the maintenance of quality standards in real
time. Computer vision is a newtechnologywhichacquires an
image of a scene and analyses it using computers in order to
control machines or to process it. It makes use of the
chromatic (colour) and geometric (shape, size, texture)
attributes that are present on the image to predict the
outcome.
The tasks involved in Computer vision are the methods to
acquire, process, analyze and understand the digital images,
and to extract high-dimensional data from those images so
as to produce the decision in either numerical or symbolic
forms. Artificial systems have been designed to act based on
the science of computer vision which extracts information
from images. The image data that can be used to extract the
necessary information can be of many forms, such as
sequence of video, multiple camera views from various
angles or data from medical scanners that are usually multi-
dimensional.
Computer vision alongwith artificial neural networksisused
to identify the type or category to which the given image of
orange belongs to.
2. LITERATURE SURVEY
Yudong Zhang and his associates have combined fitness-
scaled chaotic artificial bee colony (FSCABC) algorithm[7]
and feed forward neural network (FNN) which is a hybrid
method to classify objects. The acquired images were
processed for background removal and to obtain various
features to create feature space. Since, the obtained feature
space is large principal component analysis was performed
to reduce the dimensions of the feature space. And then,
these were fed to the FNN. The FSCABC algorithm was used
to train the weights of the neural network. This method
proved to acquire higher classification efficiency than other
genetic algorithms.
Yudong Zhang and his associates proposed another novel
fruit-classification system, with the goal ofrecognizingfruits
in a more efficient way. It employs four-step pre-processing
before the extraction of various geometric and chromatic
attributes. Principal component analysis was utilized in the
decrease of the quantity of features. Feed-forward neural
network (FNN)andBiogeography-basedoptimization(BBO)
[8] had been utilized to group the fruits from the diminished
features. This technique had higher productivity in both
precision and calculation time.

Wang S and his associates in their work proposed two novel
machine-learning based classification methods. Their work
involved the utilization of wavelet entropy (WE), principal
component analysis (PCA). And mainly, the two machine-
learning based classification methods namely feed-forward
neural network (FNN) along with fitness-scaled chaotic
artificial bee colony (FSCABC) and biogeography-based
optimization (BBO) were employed. This required the
utilization of considerably less features in expectationwhen
contrasted and different techniques.
Nasirahmadi and his associate Ashtiani S in their paperhave
grouped fruits and classified them using Image processing
combined with pattern recognition. They classified 20
varieties almonds which are of sweet and bitter taste
utilizing Bag-of-Feature (BOF) model. Harris, Harrise
Laplace, Hessian, Hessiane Laplace and Maximally Stable
Extremely Regions (MSER) key point indicatorsalongwitha
Scale Invariant Feature Transform (SIFT) descriptor were
utilized. The execution of Chi-SVM classifier was found be
exceptional than the k-NN and L-SVM classifiers.
Sofu MM, Erb O, Kayacan MC, Cetissli B. The paper proposed
real-time processing system that consequently sorts apples
and assesses it's quality. The apples were arranged
dependent on various characteristics, for example, colour,
size and weight into different classes. It was likewise ready
to recognize the apples that were influenced by scab and
stain and furthermore the one's that were spoiled. The
proposed framework had the capacity to sort 15 apples in a
normal for each second. It had a arranging precision rate of
73– 96%.
3. ARTIFICIAL NEURAL NETWORKS
Artificial neural network (ANN) has also been known as
connectionist system. Normally, artificial neurons are
collected into layers. It is essentiallya registeringframework
that was motivated by the working of the natural neural
systems of creature minds [1]. The neural network is
essentially not an algorithm. It is a system for the AI
calculations to chip away at to process complex information
inputs [2]. They figure out how to play out the assignments
while handling the issue by considering the precedents as
the procedure proceeds onward with no errand explicit
tenets. The process takes place without any prior
knowledge. They do so by generating characteristics that
identify the object at the time of processing.
An artificial neural network is basically a system that
consists of components called artificial neurons, which get
input, change their inward state (initiation) as indicated by
that input, and produce the necessary output upon the
information on the input and activation.
As in general, in an ANN the signs between artificial neurons
are genuine numbers, and the yield of each neuronisfigured
by some non-direct capacity of the entirety of its sources of
info. The associations between neurons are called 'edges'.
Artificial neurons and edges commonly have a weight that
alters as learning continues. The weight increments or
diminishes the quality of the flag at an association. The flag
might contain a threshold such that the signals are sent only
if the total of the flag crosses that threshold.
Typically, artificial neurons are accumulated into layers.
Distinctive layers may perform various types of changes on
their data sources. Signs travel from the primary layer (the
information/input layer), to the last layer (the yield/output
layer), conceivably in the wake of navigating the layers on
numerous occasions as in Figure 1.
Figure 1- Information flow in ANN
The network is framed by associating the yield of specific
neurons to the contribution of different neurons shaping a
coordinated, weighted diagram. The weights just as the
functions that figure the activities can be adjusted by a
procedure called learning which is administered by a
learning rule.
The learning rule is a standard or a calculation which
changes the parameters of the neural network, all together
for an offered contribution to the system to deliver a
favoured yield. This learning procedure ordinarily sums to
altering the weights and threshold of the factors inside the
system.
3.1 Components of Artificial Neural Network
Neurons
A neuron that is labelled as after receiving an input (t)
from antecedent neurons comprises of the accompanying
segments:
 an activation (t) representing the state of neuron,
 threshold , which is usually fixed,
 an activation function used in the computation of
the new activation at a time from (t), and
the (t) which gives the following relation
(t+1)= (t), (t), ,

 and an output function
(t)= (t) .
Generally, the output function is essentially the Identity
function.
Connections, weights and biases
The network comprises of associations (connections),every
association exchanging the output of a neuron to the input
of a neuron . In such a way that input neuron is the
ancestor of output neuron . Every association is appointed
a weight . Now and again an inclination term is added to
the all out weighted aggregate of contributions to fill in as a
limit to move the activation function.
Propagation function
The propagation registers the contribution to the inputs
(t) from the outputs (t) of antecedent neurons and
commonly has the structure
At the point when a bias is included with the capacity, the
above structure changes to
, where is a bias.
4. METAHEURISTICS
In software engineering and mathematical optimization, a
metaheuristic is a more elevated amount methodology or
heuristic intended to discover, create, or select a heuristic
(partial search algorithm) that may give an adequately
decent answer for an advancement issue, particularly with
fragmented or flawed data or constrained calculation limit
[5]. Metaheuristics test a lot of arrangements which is too
vast to even think about being totally inspected.
Metaheuristics in general makes a couple of presumptions
about the improvement issue being illuminated, thus they
might be usable for a variety of issues.
Contrasted with improvement calculations and iterative
techniques, metaheuristics don't ensure that an all around
ideal arrangement can be found on some class of
problems.[6] Many metaheuristics actualize some type of
stochastic advancement, so the arrangementfoundisreliant
on the arrangement of arbitrary factors generated.[5] In
combinatorial optimization, via looking over a huge
arrangement of possible arrangements, metaheuristics can
frequently discover great arrangements with less
computational exertion than enhancement calculations,
iterative strategies, or basic heuristics.[6] As such, they are
helpful methodologies for advancement issues.
Properties
These are properties that describe generallymetaheuristics:
 Metaheuristics are systems that direct the hunt
(search) procedure.
 The objective is to effectively investigate the hunt
space so as to discover near– ideal arrangements.
 Techniqueswhichcomprise metaheuristiccalculations
extend from basic neighbourhood seek techniques to
complex learning forms.
 Metaheuristic calculations are rough and more often
than not non-deterministic.
 Metaheuristics are not issue explicit.
Types of metaheuristics
There are wide assortments of metaheuristics and various
properties as for which to arrange them.
 Local search vs. global search
 Single-solution vs. population-based
 Hybridization and memetic algorithms
 Parallel metaheuristics
 Nature-inspired metaheuristics
Applications
 used for combinatorial advancement in whichanideal
arrangement is looked for over a discrete hunt space.
 multidimensional combinatorial issues, incorporating
most structure issues in designing which experience
the ill effects of the curse of dimensionality.
 Metaheuristics are additionally generally utilized for
job shop booking and occupation determination
issues.
Heuristics is a path by trial and mistaketocreatesatisfactory
answers for a perplexing issue in a sensibly down to earth
time. The point is to discover greatdoablearrangementinan
adequate timescale.
Two noteworthy parts of metaheuristic calculations are:
 Intensification
 Diversification
Diversification intends to produce differingarrangementsin
order to investigate the hunt space on the global scale, while
intensification intends to concentrate on the inquiry in a
nearby district by using the data that a present decent
arrangement is found in this area. The greatmixof thesetwo
noteworthy parts will for the most part guarantee that the
worldwide optimality.

4.1 NATURE-INSPIRED METAHEURISTICS
ALGORITHMS
Nature roused metaheuristic calculations are notable
efficient methodologies for taking care of a few hard
advancement issues. These calculations are roused from
explicit qualities of creatures existing in nature and natural
marvels.
Nature enlivened metaheuristic calculationscanbearranged
into four noteworthy divisions:
 Evolution- Based Method
 Physics-Based Method
 Swarm-Based Method and
 Human-Based Method
A portion of the nature-inspired algorithms are as per the
following:
 Ant colony optimization
 Ant lion optimizer
 Artificial bee colony algorithm
 Bat algorithm
 Cat swarm optimization
 Cuckoo search algorithm
 Differential evolution
 Firefly algorithm
 Harmony search
 Particle swarm optimization
 Shuffled complex evolution
 Simulated annealing
4.2 PARTICLE SWARM OPTIMISATION
In computational science, particleswarmoptimization(PSO)
is a strategy that upgrades an issue by iterativelyattempting
to improve a hopeful arrangement concerning a given
proportion of value. It takes care of an issue by having a
populace of candidate solutions, particles, and movingthese
particles around in the pursuit space as indicated by
straightforward numerical formulae over the molecule's
position and velocity [4].
PSO is instated with a gathering of unpredictable particles
(solutions) and then checks for optima by reviving the
generations. In each cycle, every particle is refreshed by
following two "best" values. The first is the best solution
(fitness) it has accomplished up until this point. Thisvalue is
known to be pbest. The next "best" that's taken is the best
value, obtained so distant by any particle within the
populace. This best esteem is a global best and called gbest.
lbest is chosen as the nearby best value when a particle
removes a portion of the populace neighbours.
In the wake of finding the two best values, the particle
refreshes it’s velocity and position using the equations (a)
and (b).
(a)
(b)
,the velocity of the particle, the present
particle (solution). , random number between (0,1).
, are the learning factors. Typically = =2
The decision of PSO parameters can largelyaffectoptimizing
execution. Choosing PSO parameters that yield great
execution has always been the subject of much research.
The PSO parameters can likewise be tuned by utilizing
another overlaying analyzer, an idea known as meta-
optimization, or even adjusted amid the improvement, e.g.,
by methods for fuzzy logic [4].
Advantages
 Simple to actualize.
 Just couple of parameters ought to be balanced.
4.3 USE OF METAHEURISTICS IN ORANGE
IDENTIFICATION
So as to distinguish the kinds of oranges, a fewhighlights,for
example, texture, colour, shape were recognized. The
quantity of highlights range to hundreds. Each object found
in the pictures is spoken to by a tuple of 263 qualities,
portraying texture, colour, shape and state of the comparing
object. This high dimensionality could prompt an issue of
over fitting of the classifiers to the test information, known
as the "curse of dimensionality". Anothercritical perspective
to consider is the computational proficiency of the
procedure, since the count of all the 263 highlights could be
a container neck to accomplish continuous preparing.
Consequently, a few creatorshaveproposeddistinctiveways
to deal with selecting the best highlights from a lot of
accessible factors. Thefundamental ruleofthesestrategiesis
a blend of artificial neural networks (ANN) and a nature-
inspired metaheuristic algorithm.The motivationbehindthe
ANN is to give a classifier to the information tuples, which
comprises in an established multilayer perception (MLP).
The job of the second calculation is to control the
progressive executions of the ANN, in an iterative way, until
the ideal parameters of the system are chosen. The
metaheuristic calculations were utilized in the
investigations, offering ascend to half breed highlight
determination approaches. Here we use Particle Swarm
Optimization to train the neural network.

4.4 ARTIFICIAL NEURAL NETWORK – PARTICLE
SWARM OPTIMISATION
Classification is the last stage in typical computer vision
systems, such as those intended for fruit varietyrecognition.
A set of tuple of features is given as input, the classifier is to
predict the most likely class to which the given input tuple
belongs to. In order to perform the task, the available data
are divided and grouped into training and testing groups.
The first part is used in training of the classifier, while the
later is used to evaluate it's precision. In these cases, 60% of
the samples of the three classes are randomly selected for
the process of training and validation, and the remaining
40% of the samples are included in the testing group.
The basic classifier in the proposed framework is a MLP
neural network. MLPs are characterized by a progression of
parameters which decide the type of the system: number of
layers; number of neurons in each layer; transfer function in
each layer; back-proliferation training sequence; and back
propogation weight/inclination learning function. It is
notable that proper tuning of these parameters can
fundamentally influence the precision of the technique. To
the component selection procedure, a hybrid approach is
proposed to locate the ideal parameters of the MLP. The
thought comprises in utilizing a metaheuristic calculation,
which executes the MLP with various setups of the
parameters.
PSO is a promising strategy to prepare ANN. It isquickerand
shows signs of improvement results as a rule. It helps in
computing the system loads to locate the best reasonable
answer for the given issue. PSO indicates quicker assembly.
Particle swarm optimisation's outcomes present systems
with great speculation on the informational indexes.
Contrasted with evolutionary algorithms, PSO is quicker at
drawing near to optima, however it can't adjust its speed
step sizes for fine tuning.
5. IMPLEMENTATION
Figure 2 - Block diagram of the proposed method for
orange identification.
Figure 2 shows the block diagram of the proposed method
for orange identification. A total of 12 input images 4 for
each type of oranges are taken. All of the images were
captured with high-resolution digital camera with a
resolution of 1280 * 960 pixels.
The image is then resized into 256*256 pixels and the
resized image is displayed in Figure 3.
Figure 3 - Input images for three types of oranges
The image of the fruit is then segmented from it’s
background. Thresholding is the leastcomplextechniquefor
segmenting an image and can be used to produce binary
image from the gray scale image.Theorangesaresegregated
and have a black background and therefore the
segmentation issue is direct.
The shape of object is characterized by their boundaries
which are nothing but linked edges. Boundaries are used in
the computation of the geometrical features such as size or
orientation of the image. Here, sobel operator is used to
perform the edge detection operation. After edge detection,
binary representation is used to differentiate the image and
it’s background. Morphological operations were applied to
clean small noise areas and refine the shape of the objects.
For this open and close operators were used and the image
thus obtained is displayed in Figure 4.
Figure 4 - Binary output of the oranges after edge
detection.
Feature extraction assumes a major role in computer vision
algorithms, in light of the fact that a bigger quantities of
Input image
Pre-
processing
Feature
extraction
ANN-PSO
classificationOutput

them builds the likelihood of discovering highlights that
differences the diverse existing classes. For this reason, 263
features of three types have been found and taken into
account. Some of the features are contrast, correlation,
energy, entropy, homogeneity, Standard deviation,
normalized value, mean and variance. And then, these are
fed to the ANN-PSO network in order to identify the given
image and the results are shown in Figure 5
Figure 5 - The results of the identified orange type.
6. CONCLUSIONS
Exact and productive identification of natural products is a
fundamental issue in post-gathering processes completed in
business manufacturing plants. Programmed
acknowledgment of organic productassortmentsrepresents
an open test, because of the extraordinary comparability's
that may exist between them. In this paper, another
methodology for perceiving three regular orange
assortments has beendisplayed,exhibitingit’sdownto earth
practicality through a broad arrangement of trials. The
proposed framework can be effectively inserted into the
handling chain of a nourishment plant. To start with,natural
products are fragmentedutilizinga satisfactorythresholding
in the RGB space and mathematical morphology operators,
accomplishing an exceptionally exact discovery of the
oranges.
The incredible preferred standpoint of the proposed
methodology is that it tends to be specifically applied to
different organic products, varieties and properties, since
the choice of highlights and the setup of the system are
naturally performed by metaheuristic calculations.
7. FUTURE WORK
Different AI and soft computing strategies are to be utilized
for natural product arrangement and deformitydiscoveryto
give better quality item at the customer end for future work
centers around the advances in programmed organic
product characterization utilizing soft computing. The
present system cannot classifyoverlappedimagesofobjects.
Hence, in future, the system should be upgraded to identify
overlapped objects.
REFERENCES
[1] "Artificial Neural Networks as Models of Neural
Information Processing | Frontiers Research Topic".
Retrieved (2018).
[2] "Build with AI | DeepAI". DeepAI. (2018).
[3] Kennedy, J., Eberhart, R. "Particle Swarm
Optimization". (1995)
[4] Particle Swarm Optimization -
http://www.swarmintelligence.org/tutorials.php
[5] Bianchi, Leonora; Marco Dorigo; Luca Maria
Gambardella; Walter J. Gutjahr. "A survey on
metaheuristics for stochastic combinatorial
optimization". Natural Computing (2009).
[6] Blum, C. Roli, A. "Metaheuristics in combinatorial
optimization (2003).
[7] Zhang Y, Wang S, Ji G, Phillips P. Fruit classification
using computer vision and feed-forward neural
network. J Food Eng (2014).
[8] Zhang Y, Phillips P, Wang S, Ji G, Yang J, Wu J. “Fruit
classification by biogeography-based optimization
and feed-forward neural network.” Exp Syst (2016).
[9] Wang S, Zhang Y, Ji G, Yang J, Wu J, Wei L. “Fruit
classification by wavelet-entropy and feed-forward
neural network trained by fitness-scaled chaotic ABC
and biogeography-based optimization”. Entropy
(2015).
[10] Mohammadi V, Kheiralipour K, Ghasemi-
Varnamkhasti M. “Detecting maturity of persimmon
fruit based on image processing technique.” Sci
Hortic-Amsterdam (2015).
[11] Nasirahmadi A, Ashtiani S-HM. “Bag-of-featuremodel
for sweet and bitter almond classification”. Biosyst
Eng (2017).
[12] Sofu MM, Erb O, Kayacan MC, Cetissli B. “Design of an
automatic apple sorting system using machine
vision.” Comp Electron Agric (2016).
[13] Bonabeau E, Dorigo M, Theraulaz G. “Swarm
intelligence: from natural to artificial systems”
(1999).
[14] Ciptohadijoyo S, Litananda WS,RivaiM,PurnomoMH.
“Electronic nosebasedonpartitioncolumnintegrated
with gas sensor for fruit identification and
classification.” Computer Electron Agric (2016).
[15] Garcıá-Mateos G, Hernandez JL, Escarabajal-
Henarejos D, Jaeń-Terrones S, Molina-Martıńez JM.
“Study and comparison of color models for automatic
image analysis in irrigation management
applications.” Agric Water Manage (2015).
[16] Golzarian MR, Frick RA. “Classification of images of
wheat, ryegrass and brome grass species at early
growth stages using principal component analysis.”
Plant Meth (2011).

IRJET- Artificial Neural Network and Particle Swarm Optimization in Orange Identification

More Related Content

What's hot (19)

Similar to IRJET- Artificial Neural Network and Particle Swarm Optimization in Orange Identification (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Artificial Neural Network and Particle Swarm Optimization in Orange Identification