Using Machine Learning Techniques for Evaluating

Tomato Ripeness
Nashwa El-Bendary1,4 , Esraa El hariri2,4 , Aboul Ella Hassanien3,4 , Amr
Badr3
1
Arab Academy for Science, Technology, and Maritime Transport, Cairo, Egypt
2
Faculty of Computers and Information, Fayoum University
3
Faculty of Computers and Information, Cairo University, Egypt
4
Scientific Research Group in Egypt [Link]

Abstract
Tomato quality is one of the most important factors that helps ensuring
a consistent marketing of tomato fruit. As ripeness is the main indicator
for tomato quality from customers perspective, the determination of tomato
ripeness stages is a basic industrial concern regarding tomato production in
order to get high quality product. Automatic ripeness evaluation of tomato
is an essential research topic as it may prove benefits in ensuring optimum
yield of high quality product, this will increase the income because tomato
is one of the most important crops in the world. This article presents an
automated multi-class classification approach for tomato ripeness measure-
ment and evaluation via investigating and classifying the different matu-
rity/ripeness stages. The proposed approach uses color features for classify-
ing tomato ripeness [Link] approach proposed in this article uses Prin-
cipal Components Analysis (PCA) in addition to Support Vector Machines
(SVMs) and Linear Discriminant Analysis (LDA) algorithms for feature ex-
traction and classification, respectively. Experiments have been conducted
on a dataset of total 250 images that has been used for both training and
testing datasets with 10-fold cross validation. Experimental results showed
that the proposed classification approach has obtained ripeness classification
accuracy of 90.80%, using one-against-one (OAO) multi-class SVMs algo-

Email addresses: [Link]@[Link] (Nashwa El-Bendary1,4 ),

<[Link]@[Link]> (Esraa El hariri2,4 ), aboitcairo@[Link] (Aboul
Ella Hassanien3,4 ), [Link]@[Link] (Amr Badr3 )

Preprint submitted to Expert Systems with Applications September 29, 2014

rithm with linear kernel function, ripeness classification accuracy of 84.80%
using one-against-all (OAA) multi-class SVMs algorithm with linear kernel
function, and ripeness classification accuracy of 84% using LDA algorithm.
Keywords: image classification, features extraction, ripeness, principal
component analysis (PCA), support vector machines (SVMs), linear
discriminant analysis (LDA)

1. Introduction
Fruits and vegetables development is characterized by a short period of
cell division followed by a longer period of cell elongation by water uptake.
The final fruit size mainly depends on initial cell number, rather than cell
size Cowan (2001). Fruit ripening, on the other hand, is characterized by the
development of color, flavor, texture and aroma. The actual time from an-
thesis until full maturity can vary tremendously among species/cultivars due
to genetic and environmental differences. Even between fruit on the same
plant, fruit development and ripening can take more or less time depend-
ing on local microclimate conditions and differences in sink/source relations
within the plant. In addition, when a fruit is harvested, the time of anthesis
of a particular fruit is generally unknown, as is its full history El Hariri et al.
(2014); Lang and Hübert (2012); Wei et al. (2014).

According to (FAOSTAT Database) Food and of The United Nations

(FAO-UN), tomatoes world production was about 162 million tons fresh fruits
produced in the year 2012 with income about 592 trillion dollars. Tomatoes
are taking an important place among the fruits and vegetables all over the
world, due to their continuously prevailing daily nutrition, dietary value and
production income. Moreover, tomato is the fourth most important crop
next to soybeans at world production Camelo (2004). Tomato production
has been reported for 144 countries and Egypt occupies the fifth place in these
countries at tomato production in both income and weight of fruit produced.
Furthermore, tomato is the first most important crop in Egypt. Based on the
previously stated facts, in this article we selected to focus on monitoring the
ripeness stages of tomato crop, especially in Egypt. Another matter of fact is,
on the categorization of crops into the two types of climacteric that are able
to continue ripening after picking from the mother plant and non-climacteric
that can ripe only when it is attached to the mother plant Camelo (2004);

2
Coates and Johnson (1997), tomato belongs to climacteric category as it can
reach over-ripening stage after being harvested. Also, tomatoes have many
different ripeness stages, which are 1) Green, 2) Breaker, 3) Turning, 4)
Pink, 5) Light red and 6) Red stages, so they reach full red color even when
harvested green. Red stage is the most preferred ripeness stage commercially
El Hariri et al. (2014); Camelo (2004).
As has been noted, monitoring and controlling produce (fruits and veg-
etables) ripeness has become a very important issue in the crops industry,
since ripeness is perceived by customers as the main quality indicator. Also,
the product’s appearance is one of the most worrying issues for producers
as it has a high influence on product’s quality and consumer preferences.
However, up to this day, optimal harvest dates and prediction of storage
life are still mainly based on subjective interpretation and practical expe-
rience El Hariri et al. (2014). Hence, automation of that process is of a
great gain for agriculture and industry fields. For agriculture, it may be
used to develop automatic harvest systems and saving crops from damages
caused by environmental changes. On the other hand, for industry, it is used
to develop automatic sorting system or checking the quality of fruits to in-
crease customer satisfaction level Brezmes et al. (2000); Elhariri, El-Bendary,
Fouad, Platoš, Hassanien and Hussein (2014). Accordingly, an objective and
accurate ripeness assessment of agricultural crops is important in ensuring
optimum yield of high quality products. Moreover, identifying physiologi-
cal and harvest maturity of agricultural crops correctly, will ensure timely
harvest to avoid cutting of either under-ripe and over-ripe agricultural crops
El Hariri et al. (2014); Elhariri, El-Bendary, Fouad, Platoš, Hassanien and
Hussein (2014); May and Amaran (2011).

To put it briefly, the main research motivation of the approach proposed

in this article is providing an automated multi-class classification approach
for tomato ripeness measurement and evaluation via investigating and clas-
sifying the different maturity/ripeness stages based on the color features. As
previously stated, among the 144 countries reported for tomato production,
Egypt occupied the fifth place in both income and weight of fruit produced
as well as the fact that tomato is the first most important crop in Egypt.
Thus, another research motivation in this article is the selection to focus
on monitoring the ripeness stages of tomato crop, especially in Egypt. The
dataset used for experiments were constructed based on real sample images
for tomato at different stages, which were collected from different farms in

3
Minya city, Upper Egypt. Dataset of total 250 images was used for both
training and testing datasets with 10-fold cross-validation. Training dataset
is divided into 5 classes representing the different stages of tomato ripeness.
The proposed approach consists of three phases; namely pre-processing, fea-
ture extraction, and classification phases. During pre-processing phase, the
proposed approach resizes images to 250x250 pixels, in order to reduce their
color index, and the background of each image has been removed using back-
ground subtraction technique. Also, each image has been converted from
RGB to HSV color space. For feature extraction phase, Principal Compo-
nent Analysis (PCA) algorithm was applied in order to generate a feature
vector for each image in the dataset. Finally, for classification phase, the
proposed approach applied Support Vector Machines (SVMs) and Linear
discriminant analysis (LDA) algorithms for classification of ripeness stages.
Another basic research motivation is that, to the best of our knowledge,
none of the recent ripeness classification related research works have ad-
dressed the dependency of the classification approach performance on statis-
tics of the experimented dataset(s).
So, another contribution of this article is that it highlights the most ap-
propriate classification algorithm considering the dependency of the clas-
sification approach performance on statistics of the experimented dataset.
That has been achieved via adopting the utilization of principal component
analysis (PCA) in addition to Support Vector Machines (SVMs) and Linear
Discriminant Analysis (LDA) algorithms for feature extraction and classifi-
cation, respectively, for tomato ripeness stages evaluation and classification
considering the color features. Also, both training and testing datasets have
been generated via employing the 10-fold cross validation.
An essential finding is that the performance of LDA and SVMs was highly
dependent on statistics of the dataset. That is, on datasets with fewer classes
(ripeness categories), and many training examples per class, SVMs had an
advantage over the LDA classification approach.
The selection of both SVMs classification algorithm depended on the
facts that the application of SVMs classification algorithm has may advan-
tages such as, it deliver a unique solution, it doesn’t need any assumptions
about the functional form of the transformation, because the kernel implicitly
contains a non-linear transformation. Also, if an appropriate generalization
grade was chosen, even when the training sample has some bias, SVMs can
be robust. Moreover, by choosing an appropriate kernel, one can put more
stress on the similarity between samples. However, there is as well some lim-

4
itations for using SVMs algorithm, that is the lack of transparency of results
and the need for very large training time when using large datasets Auria and
Moro (2008). On the other hand, the selection of both LDA classification al-
gorithm depended on its advantages that are LDA has some advantages such
as, the employment of projection that solves the problem of illumination by
maximizing between-class scatter and minimizing within-class scatter and it
need less samples in order to obtain a reliable classifier. However, one com-
mon disadvantage of LDA is the singularity problem as well as it fails when
all scatter matrix are singular Kumar and Kaur (2012).
In general, the limitations we faced in this research are the dataset size
that’s needed to be larger, as the accuracy of SVMs increases by increasing
the number of images per training class, and accordingly a maximum accu-
racy of 90.2% has been achieved.

The rest of this article is organized as follows. Section 2 introduces some

recent research work related to monitoring and classification of maturity
stages for tomatoes and other fruits/vegetables. Section 3 presents the core
concepts of SVMs, LDA and PCA algorithms. Section 4 describes the dif-
ferent phases of the proposed content-based classification system; namely
pre-processing, feature extraction, and classification phases. Section 5 dis-
cusses the tested image dataset and presented the obtained experimental
results. Finally, Section 6 presents conclusions and addresses a number of
future research suggestions.

2. Related work
This section reviews a number of current research approaches that tackle
the problem of ripeness monitoring and classification for tomatoes and other
fruits/vegetables.

First of all, for tomato ripeness classification, various research works have
been proposed. In Zhang and McCarthy (2012), authors offered tomato ma-
turity evaluation approach using magnetic resonance imaging (MRI). For the
proposed approach, MR images were captured for tomatoes that were har-
vested from the field at different maturity stages. Then, for each of the MR
images, the mean and histogram features of the voxel intensities in the region
of interest (RoI) were calculated. Finally, partial least square discriminant
analysis (PLS-DA) algorithm was applied using both the calculated features

5
and maturity classes variables in order to deduce a maturity classification
model showing that different maturity stages are embedded in MR images
signal intensity.
Also, in Baltazar et al. (2008), authors used 128 tomato samples that were
harvested and preliminarily sorted with colorimeter choosing only those with
roughly breaker color, which represent the ripeness stage where there is a def-
inite break in color from green to tannish-yellow. So, they firstly applied data
fusion to nondestructive image of fresh intact tomatoes by assessing both of
colorimeter and nondestructive firmness measurements for the samples at the
selected testing days using two sensors placed at different points. Then, the
measurements data were normalized. Finally, a three-class Bayesian classifier
was applied and the results showed that multi-sensorial data fusion is better
than single sensor data and considerably reduces the classification error.
Moreover, in Polder et al. (2002), authors proposed an approach based on
spectral images analysis to measure the ripeness of tomatoes for automatic
sorting. The proposed approach compared hyper-spectral images with stan-
dard RGB images for classification of tomato ripeness stages. The proposed
classification approach based on individual pixels and includes gray reference
in each image for obtaining automatic compensation of different light sources.
The proposed approach in Polder et al. (2002) applied the linear discriminant
analysis (LDA) algorithm as a classification technique depending on pixels
values and proved that spectral images are better than standard RGB images
for measuring ripeness stages of tomatoes via offering more discriminating
power.

On the other hand, for Oil palm ripeness classification, in Fadilah and
Mohamad-Saleh (2014), authors proposed an automated ripeness classifica-
tion system based on color feature for the problem of oil palm fresh bunch
ripeness classification. They used the color of oil palm fresh bunch as a
ripeness indicator. The proposed system firstly applied image segmentation
using K-means clustering algorithm to separate fruits pixels from spikes ones.
Then, it extracted a hue histogram of 100 bins for each image as feature vector
via applying two different techniques; namely principal component analysis
(PCA) and stepwise discriminant analysis (SDA), for color features reduction
purpose. Finally, it applied Artificial Neural Network (ANN) as classifier to
classify ripeness into four categories: un-ripe, under-ripe, ripe and over-ripe.
Results showed that reducing the color features using stepwise discriminant
analysis improved the performance of classification accuracy by more than

6
10%.
Also, in Bensaeed et al. (2014), authors proposed a hyperspectral-based
classification system for the purpose of classifying the ripeness of oil palm
fresh fruit bunches (FFBs). A dataset of total of 469 fruits for three types of
oil palm FFBs (nigrescens, virescens, oleifera) has been collected from MPOB
farm area at Kluang, Johor, Malaysia to be classified into three ripeness cat-
egories: over-ripe, ripe, and under-ripe. The proposed system firstly scanned
oil palm FFBs using a hyperspectral device. Then, a pixel spectral process-
ing step was performed by applying, background removal, followed by pixel
discrimination (only reflectance data was analyzed). A Low Pass Filter was
applied to data for noise reduction purpose. Finally, it applied artificial neu-
ral network (ANN) as a classifier to classify the different wavelength regions
on oil palm fruit through pixel-wise processing. This system achieved an
accuracy of more than 95% for all three types of oil palm fruits.
Moreover, in Jaffar et al. (2009), authors applied photogrammetric method-
ology in order to depict a relationship between the color of the palm oil fruits
and their ripeness stages, then they have been sorted out physically. That
proposed approach was considered as the first automation form of palm oil
grading systems. Other previous grading systems faced difficulty of using
the average color digital number values at RGB color space for determining
ripeness, due to the fusion of palm fruit images with dirt and branches. The
proposed approach applied the K-means clustering and segmented the fruit
fresh bunches (FFBs) colors in an automated fashion. Then, to differentiate
ripe FFBs from unripe fruits, the computed color value to R/G and R/B
ratios of the digital number of the segmented images was utilized.
Also, in May and Amaran (2011), authors proposed an assessment ap-
proach for ensuring optimum yield of high quality oil in order to overcome
subjectivity and inconsistency of manual human grading techniques based
on experience. Palm ripeness stages were classified into under-ripe, ripe and
over-ripe depending on different color intensity. The developed approach is
an automated ripeness assessment using RGB and fuzzy logic feature extrac-
tion and classification model to assess the ripeness of oil palm. It depended
on color intensity and achieved an accuracy of 88.74%.
Furthermore, in Zhang et al. (2014), authors proposed a method for clas-
sifying harvested dates according to their color. After images capturing,
a threshold segmentation method was applied to separate fruit from back-
ground. From RGB color space, they used only R-G plane, because blue
channel does not give effective information for dates grading. For train-

7
ing stage, to generate 2D histogram (one for each maturity class), the co-
occurrence of every color in R and G channel in each class was counted. After
2D histogram creation, it was normalized. Then, back projection matrix was
generated. For grading stage, After background removal and extraction of R
and G values, a back projection step followed by color index analysis were
performed, then color grading was computed using some statistics. The pro-
posed method in Zhang et al. (2014) achieved good results in addition to not
requiring complicated training process and machine learning algorithms.

Moreover, for Hass Avocados ripeness classification, in Guerrero and Be-

navides (2014), authors proposed an automated classification system based
on Fishers Linear Discriminant Analysis and K-means algorithms. The RGB
color space has been used and the proposed system applied some filters to
minimize noises. Then, an image segmentation step was applied, using Fish-
ers Linear Discriminant Analysis algorithm, to separate Avocado fruit from
background. Finally, K-means grouping technique was employed, in order to
classify Avocados into very mature avocados from mature and green avoca-
dos categories, based on pixels percentage. The proposed system achieved
an accuracy of 87.85%.

While for banana ripeness classification, a system based on artificial neu-

ral network (ANN) with image processing approach for color recognition has
been designed in Paulraj et al. (2009) for identifying the ripeness stage of
bananas. The proposed system depended on RGB color components of the
captured images of banana. It used four sets of bananas with different sizes
and ripeness stages. Each image of the banana was captured in four different
positions and the images were captured daily until all bananas turned to be
rotten. The proposed research used supervised Neural Network model with
utilizing the error back propagation model. It achieved an identification ac-
curacy of 96%.

Also, for watermelon ripeness classification, a system based on artifi-

cial neural network (ANN) with image processing approach was designed in
Shah Rizam et al. (2009). The proposed system measured and determined
the ripeness and quality of watermelon based on its colors in YCbCr Color
Space. It measured the ripeness by checking textured founded on the skin
of watermelon, which was classified into ripeness index from the segmented
image depending on the amount of pixels at each region. The proposed re-

8
search approach achieved an accuracy of 86.51%.

For jatropha ripeness classification, in Syal et al. (2013), authors pro-

posed an approach based on image processing techniques and fuzzy logic for
fruit sorting and grading. The proposed jatropha fruit grading system based
on three basic features: RGB color components, shape, and size of the fruit
object. In the proposed system, there are three grades ( A, B, C) for jat-
ropha fruit ripeness depending on the selected features values. The selected
features are extracted using image processing techniques, then fuzzy system
was applied to decide the grade of each fruit. Experimental results showed
that using fuzzy logic in the proposed system achieved accurate and very
promising classification.

Furthermore, there are many research works regarding some other fruits.
In Dadwal and Banga (2012), authors proposed an approach based on color
image segmentation and fuzzy logic technique to classify apples into ripe,
under-ripe and over-ripe categories. The proposed approach based on RGB
color components, where for each fruit four images were captured from dif-
ferent directions. Then, segmentation algorithm was applied to these images
for getting area of interest and the mean value for each color component
(R, G and B) was calculated. Finally, the fuzzy logic approach was applied
to decide the category of apple ripeness depending on mean values of Red,
Green and Blue color components.
In Dolaty (2012), authors proposed an approach based on image processing
for cherries sorting and grading. The proposed approach depended on the
RGB color components of the captured cherry images. It used cherry samples
from four different stages that were collected with an interval of 5 days. The
sorting system of cherries according to their ripeness used color criteria and
the Total Soluble Solids (TTS) in fruit to classify it into the right ripeness
stage. In order to minimize the error rate in calculating the average color
components, the proposed system achieved 92% accuracy in sorting cherries
according to their ripeness stage.

In Damiri and Slamet (2012), authors proposed an approach based on

image processing and Artificial Neural Network (ANN) to identify maturity
and ripeness of lime. The proposed system depended on area, shape factor,
RGB color index, and texture features to identify the ripeness of lime. These
features were sent to the ANN using back propagation method for training

9
then to implement classification. The proposed system achieved 100% accu-
racy in classifying the lime based on their maturity and ripeness.

Also, in one of our previous researches Elhariri, El-Bendary, Hussein, Has-

sanien and Badr (2014), we proposed an automated ripeness classification
system for bell pepper ripeness stages. The used datasets in that research
were constructed based on real sample images for tomato at different stages,
which were collected from different farms in Minya city, Upper Egypt and
annotated by an expert. The proposed system firstly employs preprocessing
step by applying image resizing, background removal and converting image
from RGB to HSV color space. Then, 1D(16 × 4 × 4) HSV colored histogram
and color moments were computed to be used for classification purpose. Prin-
cipal component analysis (PCA) algorithm was applied as feature extraction
in order to generate a feature vector for each image in the dataset. Then,
One-against-one Support Vector Machines (SVMs) algorithm was applied for
classification of ripeness stages. The proposed system achieved an accuracy
of 93.89%.

So, on the whole, many points of research concerning ripeness classifica-

tion for tomatoes and other fruits/vegetables have been addressed. Most of
those approaches used various classifiers, such as ANN, K-means, LDA, and
fuzzy logic approaches, and reported the obtained classification accuracy of
the implemented classification approach. However none of them highlighted
the dependency of the classification approach performance on statistics of
the experimented dataset(s). So, to the best of our knowledge, this article
is the first research work aims at highlighting the most appropriate classifi-
cation algorithm considering the dependency of the classification approach
performance on statistics of the experimented dataset. Accordingly, in this
article we adopted the utilization of principal component analysis (PCA) in
addition to Support Vector Machines (SVMs) and Linear Discriminant Anal-
ysis (LDA) algorithms for feature extraction and classification, respectively,
for tomato ripeness stages evaluation and classification considering the color
features. In the proposed approach, we have conducted experiments on a
dataset of total 250 real sample images for tomato at different stages that
was used for both training and testing datasets with 10-fold cross validation.
We figured out that the performance of LDA and SVMs was highly depen-
dent on statistics of the dataset. That is, on datasets with fewer classes
(ripeness categories), and many training examples per class, SVMs had an

10
advantage over the LDA classification approach. Therefore, the obtained
experimental results showed that using one-against-one (OAO) multi-class
SVMs algorithm with linear kernel function outperformed using both one-
against-all (OAA) multi-class SVMs algorithm with linear kernel function
and LDA classification algorithms.

3. Preliminaries
This section presents a brief idea concerning the core concepts of PCA
feature extraction algorithm in addition to SVMs and LDA classification al-
gorithms.

3.1. Principal component analysis (PCA)

One of the most common statistical techniques is Principal component
analysis (PCA). It is widely used in fields of compression for a dimensionality
reduction and image recognition , data representation and features extraction
tool as it guarantees better classification Suganthy and Ramamoorthy (2012);
Ada and Kaur (2013); El-Bendary et al. (2011); Xiao (2010). Basically, It
reduces the dimensionality by avoiding redundant information, and trans-
forming samples features space to features sub-space (smaller space which
contains all independent variables which are needed to describe the data)
via disposing all ineffectual minor components. So, it’s very important to
perform many pre-processing steps in order to utilize the PCA method for
feature extraction. Steps of PCA algorithm are shown in algorithm (1).

3.2. Color Features

A widely used feature in image retrieval and image classification problems
is the color, which is as well an important feature for image representation
Shahbahrami and Juurlink (2008). In this research two color descriptors are
used; namely color moments and color histogram.

3.2.1. Color moments

The first three color moments, which are mean, standard deviation, and
skewness Soman et al. (2012); Singh and Hemachandran (2012), have been
proved to be efficient and effective way for representing color distribution in
any image. Mean, standard deviation, and skewness for a colored image of
size N × M pixels are defined by equations (1), (2), and (3).

11
Algorithm 1 Principal component analysis (PCA) Algorithm
1: Calculate the sample mean µ̄
Pn
Xi
µ̄ = i=1
n
2: Subtract sample mean from each observation Xi

Z̄i = Xi − µ̄

3: Calculate the covariance matrix C

n
X
C= Zi · Zit
i=1

4: Calculate the eigenvectors and eigenvalues of the covariance matrix C

5: Rearrange the eigenvectors and eigenvalues and select a subset as basis
vectors
6: Project the data

PM ·N
xij
j=1
x̄i = (1)
M ·N
v
u M.N
u 1 X
∂i = t (xij − x̄i )2 (2)
M · N j=1
v
u M.N
u 1 X
Si = t
3
(xij − x̄i )3 (3)
M · N j=1

Where, xij is the value of image pixel j of color channel i (e.g RGB,
HSV and etc..), x̄i is the mean for each channel i=(H,S and V), ∂i is the
standard deviation, and Si is the skewness for each channel Soman et al.
(2012); Singh and Hemachandran (2012). HSV channels can be computed
for RGB channels using equations (4), (5), and (6), where R, G and B are
color component of RGB color space Meskaldji K (2009).

12
 1
−1 2
[(R
− G) + (R − B)]
H = cos p (4)
(R − G)2 + (R − B)(G − B)
3[min (R, G, B)]
S =1− (5)
(R + G + B)
R+G+B
V =( ) (6)
3
3.2.2. Color histogram
Color histogram is a color descriptor that shows representation of the
distribution of colors in an image. It represents the number of pixels that
have colors in each range of colors El-Bendary et al. (2011). Color histogram
can be calculated for many color spaces (e.g. RGB, HSV, etc). It is often used
with 3-dimensional spaces like as RGB and HSV color spaces. color histogram
is invariant with rotation, translation, and scale Meskaldji K (2009).

3.3. Support Vector Machines (SVMs)

3.3.1. Basics
The Support Vector Machines (SVMs) is a Machine Learning (ML) al-
gorithm that is used for classification and regression of high dimensional
datasets with great results Wu and Zhou (2006); Zawbaa, El-Bendary, Abra-
ham and Hassanien (2011); Zawbaa, El-Bendary, Hassanien and Kim (2011).
SVMs algorithm solves the classification problem via trying to find an opti-
mal separating hyperplane between classes. it depends on the training cases
which are placed on the edge of class descriptor this is called support vec-
tors, any other cases are discarded Tzotsos and Argialas (2006); Zhang et al.
(2010); Suralkar et al. (2012); El Hariri et al. (2014).
SVMs algorithm seeks to maximize the margin around a hyperplane that
separates a positive class from a negative class Wu and Zhou (2006); Zawbaa,
El-Bendary, Abraham and Hassanien (2011); Zawbaa, El-Bendary, Hassanien
and Kim (2011); El Hariri et al. (2014). Given a training dataset with n
samples (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ), where xi is a feature vector in a v-
dimensional feature space and with labels yi ∈ −1, 1 belonging to either of
two linearly separable classes C1 and C2 . Geometrically, the SVMs modeling
algorithm finds an optimal hyperplane with the maximal margin to separate
two classes, which requires to solve the optimization problem, as shown in
equations (7) and (8).

13
 n n
X 1X
maximize αi − αi αj yi yj .K(xi , xj ) (7)
i=1
2 i,j=1
n
X
Subject − to : α i yi , 0 ≤ α i ≤ C (8)
i=1

Where, αi is the weight assigned to the training sample xi . If αi > 0, xi

is called a support vector. C is a regulation parameter used to trade-off the
training accuracy and the model complexity so that a superior generalization
capability can be achieved. K is a kernel function, which is used to measure
the similarity between two samples. Different choices of kernel functions
have been proposed and extensively used in the past and the most popular
are the gaussian radial basis function (RBF), polynomial of a given degree,
linear, and multi layer perceptron MLP. These kernels are in general used,
independently of the problem, for both discrete and continuous data.

3.3.2. N-class Support Vector Machine Approaches

There are many problem, which are N-class classification problem not
a binary one and it is well known that SVMs algorithm is a binary class
classification methods SVMs algorithm is a binary class classification method
and our problem is a N-class classification problem. Therefore, to use SVMs
for N-class classification problem like our problem in this article, there are
two different approaches can be applied in order to do that; namely one-
against-all (OAA) and one-against-one (OAO) approaches Liu and Zheng
(2005); Anthony et al. (2007).
The first approach, one-against-all (OAA), worked according to Algo-
rithm (2).

Algorithm 2 One-against-all (OAA)

1: Construct N binary SVMs.
2: Each SVM separates one class from the rest classes.
3: Train the ith SVM with all training samples of the ith class with positive
labels, and training samples of other classes with negative labels.

In the second approach, one-against-one (OAO), a SVM classifier was

created for each pair of classes (forN −class problem) resulting in N (N −1)/2
classifiers. The OAO approach worked according to Algorithm (3).

14
Algorithm 3 one-against-one (OAO)
1: CreateN (N − 1)/2 binary SVMs
2: Train N (N − 1)/2 binary SVMs as follow
(1, 2), (1, 3), ..., (1, k), (2, 3), (2, 4), ..., (k1, k).
3: Select the one with the largest vote(The class label that occurs the most).

3.4. Linear Discriminant Analysis (LDA)

Linear Discriminant Analysis (LDA) is a commonly used technique for
data classification and dimensionality reduction. Linear Discriminant Anal-
ysis easily handles the case where the within class frequencies are unequal
and their performances has been examined on randomly generated test data.
it’s basic idea is to find a linear transformation that best discriminate among
classes, then classification can be performed in transformed space based on
some metrics(Euclidean distance) Li et al. (2006).
Given data yi , xi , i =n1 , where yi 1, 2, ...,K is the class label, k is the
number of classes and xi is a vector of features or predictors, we seek to find
the best direction in the predictor space in which the classes are separated
as much as possible.
Mathematically, LDA implementation is carried out via scatter matrix
analysis. For all samples of all classes, we define two measures Li et al.
(2006); Fisher (1936):

• Within-class scatter matrix. It is defined by:

Nj
K X
X
Sw = (xji − µj )(xji − µj )T (9)
j=1 i=1

Where xi j is the ith sample of class j, µj is the mean of class j, K is

the number of classes, and Nj is the number of samples in classj; and

• Between-class scatter matrix. It defined by:

K
X
Sb = (µj − µ)(µj − µ)T (10)
j=1

Where µ represents the mean of all classes. This method maximizes

the ratio of between-class measure to the within-class measure in any

15
 particular data set thereby guaranteeing maximal separability. The
δ|Sb |
maximization of δ|S w|
.

4. The Proposed System

In particular, the proposed system is capable of classifying different ripeness
stages of tomato. The proposed system exploits color features to identify
different ripeness stages of tomato in multi-class scenario. Proposed system
general structure is described at Fig. 1. It consists of three phases; namely
pre-processing, feature extraction, and classification phases.

Figure 1: Architecture of the proposed ripeness classification approach

4.1. Pre-processing phase

During pre-processing phase, In order to reduce images color index,the
proposed system resizes images to 250x250 pixels. Then the background of
each image is removed using background subtraction technique with some
morphological operations. Fig. 2. shows an example of background removal
algorithm results. Also, each image is converted from RGB to HSV color
space, as it is widely used in the field of color vision and close to the categories
of human color perception Yu et al. (2002).

16
4.2. Feature extraction phase
As previously stated, the most important characteristic to assess tomato
ripeness is its surface color, so this system uses HSV color histogram and color
moments for ripeness stages classification. For feature extraction phase, PCA
algorithm is applied as features extraction technique in order to generate a
feature vector for each image in the dataset.

Figure 2: sample images before and after applying background subtraction algorithm

The proposed system transforms the input space into sub-spaces for di-
mensionality reduction. After completing the previous 1D 16x4x4 HSV his-
togram, 16 levels for hue and 4 levels for each of saturation and value are
resulted. In addition, nine color moments, three for each channel (H, S and V
channels) (mean, standard deviation, and skewness), were computed. Then,
a feature vector was formed as a combination of HSV 1D histogram and the
nine color moments.

4.3. Classification phase

Finally, for classification phase, the proposed system applied SVMs and
LDA algorithms for classification of ripeness stages. The inputs are training
dataset feature vectors and their corresponding classes, whereas the outputs
are the ripeness stage of each image in the testing dataset.

In this phase, the classification system, previously proposed in Elhariri,

El-Bendary, Fouad, Platoš, Hassanien and Hussein (2014), has been utilized

17
along with the one-against-one (OAO) approach with 10-fold cross valida-
tion for multi-class SVMs problems. In addition to, another classification
algorithm LDA using quadratic discriminant function with 10-fold cross val-
idation.

5. Experimental Results
Simulation experiments in this article are done on a PC with Intel Core
i7 Q720 @ 1.60 GHZ CPU and 6GB memory. The proposed approach is
designed with Matlab running on Windows 7. The datasets used for experi-
ments were constructed based on real sample images for tomato at different
ripeness stages, which were collected from different farms in Minya city. The
collected datasets contained colored JPEG images of resolution 3664x2748
pixels that were captured using Kodak C1013 digital camera of 10.3 megapix-
els resolution. The dataset is of total 250 images were used for both training
and testing datasets with 10-fold cross-validation. Training dataset is di-
vided into 5 classes representing the different stages of tomato ripeness as
shown in Fig. 3. The classes are Green & Breaker, Turning, Pink, Light Red,
and Red stages U.S. Dept. Agric./AMS (1991). For Green & Breaker stage,
green represents the ripeness stage where fruit surface is completely green,
however breaker represents the ripeness stage where there is a definite break
in color from green to tannish-yellow, pink or red on not more than 10% of
the surface. For Turning stage, 10% to 30% of the surface is not green. For
Pink stage, 30% to 60% of the surface is not green. For Light Red stage, 60%
to 90% of the surface is not green. Finally, for Red stage, more than 90% of
the surface is not green. Some samples of both training and testing datasets
are shown in Fig. 4.
The proposed approach has been implemented considering three scenar-
ios; namely

• Scenario I: One-against-One multi-class SVMs system using10-fold cross

validation

• Scenario II: One-against-All multi-class SVMs system using10-fold cross

validation

• Scenario III: LDA system using10-fold cross validation

18
 Figure 3: Examples of tomato ripeness stages

Figure 4: Examples of training and testing samples

5.1. Scenario I: One-against-One multi-class SVMs system using10-fold cross

validation
The first scenario presents implementing One-against-One multi-class SVMs
system using10-fold cross-validation and a total of 250 images for both of
19
training and testing datasets. The used features for classification are a com-
bination of color HSV histogram and color moments and PCA algorithm
was applied for features extraction. Moreover, SVMs algorithm was em-
ployed with different kernel functions that are: Linear kernel, radial basis
function (RBF) kernel, and Multi-Layer Perceptron (MLP) kernel and Poly-
nomial kernel Boolchandani and Sahula (2011); Vanschoenwinkel and Man-
derick (2005) for ripeness stage classification.

Figure 5 shows classification accuracy obtained via applying each kernel

function. Figures 6, 7, 8 and 9 show 5-class receiver operating characteris-

Figure 5: Results for different SVM kernel functions using one-against-one multi-class
approach and 10-fold cross-validation

tic (ROC) curve and area under curve (AUC) for the best feature resulted
from different kernel functions using one-against-one multi-class SVMs ap-
proach with 10-fold cross-validation and total of 250 images (used for both
of training and testing). The ROC curve separates each class from other
classes. From figure 6, showing the ROC curve for the best feature using
linear kernel function for OAO multi-class SVMs with cross-validation, the

20
Figure 6: ROC curve for the best feature using linear kernel function (OAO multi-class
SVMs with cross-validation), AUC=0.8219

applied approach separated each class from each one of the rest classes by
AUCs shown at Table 1.

Table 1: AUCs of OAO multiclass-SVMs using Linear kernel function & 10-fold cross
validation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.7180 0.8615 0.9424 0.9785
Turning 0.7180 - 0.6649 0.8123 0.9268
Pink 0.8615 0.6649 - 0.6880 0.8652
Light Red 0.9424 0.8123 0.6880 - 0.7615
Red 0.9785 0.9268 0.8652 0.7615 -

From figure 7, showing ROC curve for the best feature using MLP kernel
function for OAO multi-class SVMs with cross-validation, the applied ap-
proach separated each class from each one of the rest classes by AUCs shown

21
at Table 2.

Figure 7: ROC curve for the best feature using MLP kernel function (OAO multi-class
SVMs with cross-validation), AUC=0.7217

Table 2: AUCs of OAO multiclass-SVMs using MLP kernel function & 10-fold cross vali-
dation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.6299 0.6998 0.8084 0.8775
Turning 0.6299 - 0.5985 0.7197 0.8115
Pink 0.6998 0.5985 - 0.6391 0.7521
Light Red 0.8084 0.7197 0.6391 - 0.6810
Red 0.8775 0.8115 0.7521 0.6810 -

From figure 8, showing ROC curve for the best feature using RBF kernel
function for OAO multi-class SVMs with cross-validation, the applied ap-
proach separated each class from each one of the rest classes by AUCs shown
at Table 3.

22
Figure 8: ROC curve for the best feature using RBF kernel function(OAO multi-class
SVMs with cross-validation), AUC=0.8191

Table 3: AUCs of OAO multiclass-SVMs using RBF kernel function & 10-fold cross vali-
dation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.7400 0.8646 0.9411 0.9853
Turning 0.7400 - 0.6658 0.7957 0.9231
Pink 0.8646 0.6658 - 0.6555 0.8460
Light Red 0.9411 0.7957 0.6555 - 0.7735
Red 0.9853 0.9231 0.8460 0.7735 -

From figure 9, showing ROC curve for the best feature using Polynomial
kernel function for OAO multi-class SVMs with cross-validation, the applied
approach separated each class from each one of the rest classes by AUCs
shown at Table 4.

23
Figure 9: ROC curve for the best feature using Polynomial kernel function(OAO multi-
class SVMs with cross-validation), AUC=0.8070

Table 4: AUCs of OAO multiclass-SVMs using Polynomial kernel function & 10-fold cross
validation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.8621 0.9646 0.9627 0.9234
Turning 0.8621 - 0.8093 0.7620 0.5581
Pink 0.9646 0.8093 - 0.5801 0.8601
Light Red 0.9627 0.7620 0.5801 - 0.7874
Red 0.9234 0.5581 0.8601 0.7874 -

5.2. Scenario II: One-against-All multi-class SVMs system using10-fold cross

validation
In the second scenario, the proposed One-against-All multi-class SVMs
approach was also tested using the previously stated specifications of One-
against-One multi-class SVMs approach for ripeness stages classification.
Figure 10 shows classification accuracy obtained via applying each kernel
function.

24
Figure 10: Results for different kernel functions using one-against-All multi-class approach
and 10-fold cross-validation

Figures 11,12, 13 and 14 show 5-class receiver operating characteristic

(ROC) curve and area under curve (AUC) for the best feature for different
kernel function using one-against-all multi-class SVMs approach with 10-fold
cross-validation and total of 250 images (used for both of training and test-
ing).

From figure 11, showing ROC curve for the best feature using linear
kernel function for OAA multi-class SVMs with cross-validation, the applied
approach separated each class from each one of the rest classes by AUCs
shown at Table 5.
From figure 12, showing ROC curve for the best feature using MLP ker-

25
Figure 11: ROC curve for the best feature using linear kernel function (OAA multi-class
SVMs with cross-validation), AUC=0.8154

Table 5: AUCs of OAA multiclass-SVMs using Linear kernel function & 10-fold cross
validation
Grean &Breaker - 0.9043 0.9669 0.9653 0.9284
Turning 0.9043 - 0.8259 0.8176 0.5294
Pink 0.9669 0.8259 - 0.5361 0.8436
Light Red 0.9653 0.8176 0.5361 - 0.8362
Red 0.9284 0.5294 0.8436 0.8362 -

nel function for OAA multi-class SVMs with cross-validation, the applied
approach separated each class from each one of the rest classes by AUCs
shown at Table 6.
From figure 13, showing ROC curve for the best feature using RBF ker-
nel function for OAA multi-class SVMs with cross-validation, the applied
approach separated each class from each one of the rest classes by AUCs
shown at Table 7.
From figure 14, showing ROC curve for the best feature using Polynomial
kernel function for OAA multi-class SVMs with cross-validation, the applied

26
Figure 12: ROC curve for the best feature using MLP kernel function (OAA multi-class
SVMs with cross-validation), AUC=0.7232

Table 6: AUCs of OAA multiclass-SVMs using MLP kernel function & 10-fold cross vali-
dation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.7334 0.8269 0.9089 0.7826
Turning 0.7334 - 0.6901 0.7289 0.5050
Pink 0.8269 0.6901 - 0.5110 0.7616
Light Red 0.9089 0.7289 0.5110 - 0.7836
Red 0.7826 0.5050 0.7616 0.7836 -

approach separated each class from each one of the rest classes by AUCs
shown at Table 8.

5.3. Scenario III: LDA system with 10-fold cross validation

In the third scenario, the proposed LDA multi-class system was also
tested using the previously stated specifications for ripeness stages classi-
[Link] accuracy obtained via applying LDA classifier is 84 %.

27
Figure 13: ROC curve for the best feature using RBF kernel function(OAA multi-class
SVMs with cross-validation), AUC=0.8100

Table 7: AUCs of OAA multiclass-SVMs using RBF kernel function & 10-fold cross vali-
dation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.7555 0.8565 0.9242 0.9781
Turning 0.7555 - 0.6641 0.7730 0.9242
Pink 0.8565 0.6641 - 0.6091 0.8244
Light Red 0.9242 0.7730 0.6091 - 0.7911
Red 0.9781 0.9242 0.8244 0.7911 -

Figure 15, showing ROC curve for the best feature using LDA system
with cross-validation, the applied approach separated each class from each
one of the rest classes by AUCs shown at Table 9.
From the previously depicted experimental results, we found out that the
One-against-One multi-class SVMs approach is better than the One-against-
All multi-class SVMs and LDA approaches, when applied for ripeness stage
classification. Figure 16 shows a comparison between accuracies obtained by

28
Figure 14: ROC curve for the best feature using Polynomial kernel function(OAA multi-
class SVMs with Cross-validation), AUC=0.7967

Table 8: AUCs of OAA multiclass-SVMs using Polynomial kernel function & 10-fold cross
validation
Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.8178 0.9227 0.9225 0.8771
Turning 0.8178 - 0.8480 0.8196 0.6075
Pink 0.9227 0.8480 - 0 0.5230 0.8368
Light Red 0.9225 0.8196 0.5230 - 0.7924
Red 0.8771 0.6075 0.8368 0.7924 -

each of the three approaches.

The accuracy measure is calculated as shown in equation (11).
N umberof correctlyclassif iedimages
Accuracy = ∗ 100 (11)
T otalnumberof testingimages

29
Figure 15: ROC curve for the best feature using LDA classifier with Cross-validation,
AUC=0.8531

Table 9: AUCs of LDA classifier & 10-fold cross validation

Green & Turning Pink Light Red Red
Breaker
Grean &Breaker - 0.8478 0.9621 0.9604 0.9196
Turning 0.8478 - 0.9340 0.9159 0.7055
Pink 0.9621 0.9340 - 0.5720 0.8807
Light Red 0.9604 0.9159 0.5720 - 0.8327
Red 0.9196 0.7055 0.8807 0.8327 -

6. Conclusions and Future work

In recent years, the use of computer vision and machine learning tech-
niques for evaluating/estimating fruits ripeness has become widely used.
From another perspective, tomato maturity has been related to quantifi-
able parameters, that reflect the biochemical changes during ripening. Color
is used as a major method in determining maturity of tomato. However,
skin color of tomato varies from cultivar to another cultivar even at the

30
Figure 16: Comparison between the classification accuracy of OAA, OAO multi-class
SVMs and LDA approaches

same maturity stage Molyneux et al. (2004); Zhang and McCarthy (2012);
El Hariri et al. (2014). During ripening, tomatoes go through a series of
highly ordered physiological and biochemical changes, such as chlorophyll
degradation and increased activity of cell wall-degrading enzymes, bring on
changes in color, firmness, and development of aromas and flavors El Hariri
et al. (2014); Prasanna et al. (2007). For tomatoes, ripeness issue is often
handled via classifying harvested produce according to discrete color classes
going from immature green to mature red, as stated in some recent researches
that have classified tomatoes in different maturity stages based on measure-
ments of color El Hariri et al. (2014); Hahn (2002); Aranda-Sanchez et al.

31
(2009). Different tomato products have distinct requirements for maturity
to achieve quality standards; hence, tomato maturity is one of the most im-
portant factors associated with the quality of processed tomato products.

Accordingly, automation of the ripeness stage classification process for

crops is of a great use for different domains; such as agriculture, via devel-
oping automatic harvest systems and saving crops from damages caused by
environmental changes and ensuring optimum yield of high quality products.
Also, for industry, through developing automatic sorting systems for check-
ing the quality of fruits to increase customer satisfaction level. Moreover, for
trading, identifying physiological and harvest maturity of agricultural crops
correctly, will ensure timely harvest for exporting and to avoid cutting of
either under-ripe and over-ripe agricultural crops El Hariri et al. (2014); El-
hariri, El-Bendary, Fouad, Platoš, Hassanien and Hussein (2014); May and
Amaran (2011).

In conclusion, the approach proposed in this article has one main re-
search motivation that is providing an automated multi-class classification
approach for tomato ripeness measurement and evaluation via investigating
and classifying the different maturity/ripeness stages based on the color fea-
tures. An essential finding is that the performance of classification algorithms
was highly relative to statistics of the experimented datasets. That is, on
datasets with fewer classes (ripeness categories), and many training examples
per class, SVMs had an advantage over other classification approaches. Many
points of research concerning ripeness classification for tomatoes and other
fruits/vegetables have been addressed by other researchers; however none of
those classification approaches addressed the dependency of the classification
approach performance on statistics of the experimented dataset(s).
The proposed system has three main phases; pre-processing implemented
by applying resizing, background removal, and extracting color components
for each image, PCA based feature extraction applied to each pre-processed
image in order to obtain HSV histogram and color moments feature vectors,
and finally, SVMs and LDA models are generated for ripeness stage classifica-
tion. The proposed approach has been implemented considering two scenar-
ios via applying One-against-One multi-class SVMs system, One-against-All
multi-class SVMs system and LDA system using 10-fold cross-validation.
Based on the obtained the experimental results, the highest ripeness classi-
fication accuracies of 90.80 % and 84.80% have been achieved by the first

32
scenario and the second scenarios, respectively, using linear kernel function
and 84% using third scenario. Thus, it can be concluded that the ripeness
classification accuracy obtained by the OAO multi-class SVMs approach is
better than ripeness classification accuracy obtained by the OAA multi-class
SVMs and LDA approaches.
On the other hand, the limitations we faced in this research are the dataset
size that’s needed to be larger, as the accuracy of SVMs increases by increas-
ing the number of images per training class, and accordingly a maximum
accuracy of 90.2% has been achieved using our current experimented dataset.

For future research, variety of challenges and research directions could

be considered. Some general research directions are to consider applying the
approach proposed in this article to different fields of objects classification
other than crops ripeness stages classification. For example, in industry, it
could be applied for achieving higher accuracy for objects sorting systems or
checking the quality of objects to increase customer satisfaction level. More-
over, for trading, the proposed approach can be of a great use for detecting
and choosing the best ripeness stage according to production needs such as
exporting, manufacturing, ... etc. Another open problem is to apply the pro-
posed approach on different crops, other than tomatoes, in order to automate
the whole process of harvesting and detect damages to save crops.
From the perspective of the utilized algorithms in the proposed approach,
a number of future research could be achieved via classifying different ob-
jects or crops by involving other features (texture, shape, size, ... etc.) ac-
cording to the classified objects nature. Moreover, other Machine Learning
approaches could be employed in order to address the advantages and lim-
itations of applying each of them. Another direction of research is to use
nondestructive/non-invasive detection technologies of food quality/maturity
such as hyperspectral imaging systems, colorimetric, Near Infrared Spec-
troscopy, and non-invasive smart sensing technologies.

33
References
Ada and Kaur, R. (2013). Feature extraction and principal component analy-
sis for lung cancer detection in ct scan images, International Journal of Ad-
vanced Research in Computer Science and Software Engineering 3(3): 187–
190.

Anthony, G., Gregg, H. and Tshilidzi, M. (2007). Image classification using

svms: one-against-one vs one-against-all, arXiv preprint arXiv:0711.2914
.

Aranda-Sanchez, J. I., Baltazar, A. and González-Aguilar, G. (2009). Imple-

mentation of a bayesian classifier using repeated measurements for discrim-
ination of tomato fruit ripening stages, Journal of Biosystems Engineering
102(3): 274–284.

Auria, L. and Moro, R. A. (2008). Support vector machines (svm) as a

technique for solvency analysis, Technical report, Discussion papers: Ger-
man Institute for Economic Research (Deutsches Institut fr Wirtschafts-
forschung, Berlin).

Baltazar, A., Aranda, J. I. and González-Aguilar, G. (2008). Bayesian clas-

sification of ripening stages of tomato fruit using acoustic impact and col-
orimeter sensor data, Computers and Electronics in Agriculture 60(2): 113–
121.

Bensaeed, O., Shariff, A., Mahmud, A., Shafri, H. and Alfatni, M. (2014).
Oil palm fruit grading using a hyperspectral device and machine learn-
ing algorithm, IOP Conference Series: Earth and Environmental Science,
Vol. 20, IOP Publishing, p. 012017.

Boolchandani, D. and Sahula, V. (2011). Exploring efficient kernel functions

for support vector machine based feasibility models for analog circuits,
International Journal of Design, Analysis & Tools for Circuits and Systems
1(1).

Brezmes, J., Llobet, E., Vilanova, X., Saiz, G. and Correig, X. (2000). Fruit
ripeness monitoring using an electronic nose, Sensors and Actuators B:
Chemical 69(3): 223–229.

34
Camelo, A. F. L. (2004). Manual for the preparation and sale of fruits and
vegetables: from field to market, Vol. 151, Food & Agriculture Org.

Coates, L. and Johnson, G. (1997). Postharvest diseases of fruit and vegeta-

bles, Journal of Plant Pathogens and Plant Diseases pp. 533–548.

Cowan, A. K. & Cripps, R. F. . R. E. W. . T. N. J. (2001). Fruit size: towards

an understanding of the metabolic control of fruit growth using avocado
as a model system, Physiologia Plantarum 111(2): 127–136.

Dadwal, M. and Banga, V. (2012). Estimate ripeness level of fruits using rgb
color space and fuzzy logic technique, International Journal of Engineering
and Advanced Technology 2(1): 225–229.

Damiri, D. J. and Slamet, C. (2012). Application of image processing and ar-

tificial neural networks to identify ripeness and maturity of the lime (citrus
medica), International Journal of Basic and Applied Science 1(2): 171–179.

Dolaty, H. (2012). Sorting and grading of cherries on the basis of ripeness, size
and defects by using image processing techniques, International Journal
of Agriculture and Crop Sciences(IJACS) 4(16): 1144–1149.

El-Bendary, N., Zawbaa, H. M., Hassanien, A. E. and Snasel, V. (2011). Pca-

based home videos annotation system, International Journal of Reasoning-
based Intelligent Systems 3(2): 71–79.

El Hariri, E., El-Bendary, N., Hassanien, A. E. and Badr, A. (2014). Au-

tomated ripeness assessment system of tomatoes using pca and svm tech-
niques, M. Sarfraz (Ed.), Computer Vision and Image Processing in Intel-
ligent Systems and Multimedia Technologies, IGI global, pp. 101–130.

Elhariri, E., El-Bendary, N., Fouad, M. M. M., Platoš, J., Hassanien, A. E.

and Hussein, A. M. (2014). Multi-class svm based classification approach
for tomato ripeness, Innovations in Bio-inspired Computing and Applica-
tions, Springer, pp. 175–186.

Elhariri, E., El-Bendary, N., Hussein, A. M. M., Hassanien, A. E. and Badr,

A. (2014). Bell pepper ripeness classification based on support vector
machine, The second IEEE International Conference on Engineering and
Technology (ICET 2014), IEEE.

35
Fadilah, N. and Mohamad-Saleh, J. (2014). Color feature extraction of
oil palm fresh fruit bunch image for ripeness classification, 13th Interna-
tional Conference on Applied Computer and Applied Computational Sci-
ence (ACACOS ’14), pp. 51–55.

Fisher, R. A. (1936). The use of multiple measurements in taxonomic prob-

lems, Journal of Annals of eugenics 7(2): 179–188.

Food and of The United Nations (FAO-UN), A. O. (2012). Fao statistical

yearbook 2013- world food and agriculture.
URL: [Link]

Guerrero, E. R. and Benavides, G. M. (2014). Automated system for clas-

sifying hass avocados based on image processing techniques, 2014 IEEE
Colombian Conference on Communications and Computing (COLCOM),
IEEE, pp. 1–6.

Hahn, F. (2002). Aeautomation and emerging technologies: multi-

spectral prediction of unripe tomatoes, Journal of Biosystems Engineering
81(2): 147–155.

Jaffar, A., Jaafar, R., Jamil, N., Low, C. Y. and Abdullah, B. (2009). Pho-
togrammetric grading of oil palm fresh fruit bunches, International Journal
of Mechanical and Mechatronics Engineering 9: 18–24.

Kumar, S. and Kaur, H. (2012). Face recognition techniques: Classification

and comparisons, International Journal of Information Technology and
Knowledge Management 5(2): 361–363.

Lang, C. and Hübert, T. (2012). A colour ripeness indicator for apples, Food
and Bioprocess Technology 5(8): 3244–3249.

Li, T., Zhu, S. and Ogihara, M. (2006). Using discriminant analysis for multi-
class classification: an experimental investigation, Journal of Knowledge
and information systems 10(4): 453–472.

Liu, Y. and Zheng, Y. F. (2005). One-against-all multi-class svm classification

using reliability measures, IEEE International Joint Conference on Neural
Networks(IJCNN’05), Vol. 2, IEEE, pp. 849–854.

36
May, Z. and Amaran, M. (2011). Automated ripeness assessment of oil
palm fruit using rgb and fuzzy logic technique, The 13th WSEAS In-
ternational Conference on Mathematical and Computational Methods in
Science and Engineering (MACMESE 2011), World Scientific and Engi-
neering Academy and Society (WSEAS), Stevens Point, Wisconsin, USA,
pp. 52–59.

Meskaldji K, B. S. . C. S. (2009). Color quantization and its impact on color

histogram based image retrieval accuracy, First International Conference
on Networked Digital Technologies, NDT’09, IEEE, pp. 515–517.

Molyneux, S. L., Lister, C. E. and Savage, G. P. (2004). An investigation

of the antioxidant properties and colour of glasshouse grown tomatoes,
International journal of food sciences and nutrition 55(7): 537–545.

Paulraj, M., Hema, C. R., R Pranesh, K. and Siti Sofiah, M. R. (2009).

Color recognition algorithm using a neural network model in determining
the ripeness of a banana, The International Conference on Man-Machine
Systems (ICoMMS), Universiti Malaysia Perlis, pp. 2B71–2B74.

Polder, G., Van der Heijden, G. and Young, I. (2002). Spectral image analy-
sis for measuring ripeness of tomatoes, Transactions-American Society of
Agricultural Engineers 45(4): 1155–1162.

Prasanna, V., Prabha, T. and Tharanathan, R. (2007). Fruit ripening

phenomena–an overview, Journal of Critical Reviews in Food Science and
Nutrition 47(1): 1–19.

Shah Rizam, M., Farah Yasmin, A., Ahmad Ihsan, M. and Shazana, K.
(2009). Non-destructive watermelon ripeness determination using image
processing and artificial neural network (ann), International Journal of
Electrical and Computer Engineering 4(6).

Shahbahrami, A. and Juurlink, D. B. B. (2008). Comparison between color

and texture features for image retrieval.

Singh, S. M. and Hemachandran, K. (2012). Content-based image retrieval

using color moment and gabor based image retrieval using color moment
and gabor texture teature, International Journal of Computer Science Is-
sues 9(1): 299–309.

37
Soman, S., Ghorpade, M., Sonone, V. and Chavan, S. (2012). Content
based image retrieval using advanced color and texture features, Interna-
tional Conference in Computational Intelligence (ICCIA 2012), Vol. IC-
CIA, pp. 1–5.
Suganthy, M. and Ramamoorthy, P. (2012). Principal component analysis
based feature extraction, morphological edge detection and localization for
fast iris recognition, Journal of Computer science 8(9): 1428–1433.
Suralkar, S., Karode, A. and Pawade, P. W. (2012). Texture image classifi-
cation using support vector machine, International Journal of Computer
Technology & Applications 3(1).
Syal, S., Mehta, T. and Darshni, P. (2013). Design & development of intel-
ligent system for grading of jatropha fruit by its feature value extraction
using fuzzy logics, International Journal of Advanced Research in Com-
puter Science and Software Engineering(IJARCSSE) 3(7): 1077–1081.
Tzotsos, A. and Argialas, D. (2006). A support vector machine approach
for object based image analysis, International Conference on Object-based
Image Analysis (OBIA06).
U.S. Dept. Agric./AMS, Washington, D. (1991). United states standards for
grades of fresh tomatoes.
URL: [Link]
Vanschoenwinkel, B. and Manderick, B. (2005). Appropriate kernel func-
tions for support vector machine learning with sequences of symbolic data,
Deterministic and Statistical Methods in Machine Learning, Vol. 3635,
Springer, pp. 256–280.
Wei, X., Liu, F., Qiu, Z., Shao, Y. and He, Y. (2014). Ripeness classification
of astringent persimmon using hyperspectral imaging technique, Food and
Bioprocess Technology 7(5): 1371–1380.
Wu, Q. and Zhou, D.-X. (2006). Analysis of support vector machine classi-
fication, Journal of Computational Analysis & Applications 8(2).
Xiao, B. (2010). Principal component analysis for feature extraction of im-
age sequence, 2010 International Conference on Computer and Commu-
nication Technologies in Agriculture Engineering (CCTAE), Vol. 1, IEEE,
pp. 250–253.

38
Yu, H., Li, M., Zhang, H.-J. and Feng, J. (2002). Color texture moments
for content-based image retrieval, International Conference on Image Pro-
cessing, Vol. 3, IEEE, pp. 929–932.

Zawbaa, H. M., El-Bendary, N., Abraham, A. and Hassanien, A. E.

(2011). Svm-based soccer video summarization system, 2011 Third World
Congress on Nature and Biologically Inspired Computing (NaBIC), IEEE,
pp. 7–11.

Zawbaa, H. M., El-Bendary, N., Hassanien, A. E. and Kim, T.-h. (2011).

Machine learning-based soccer video summarization system, Multimedia,
Computer Graphics and Broadcasting, Springer, pp. 19–28.

Zhang, D., Lee, D.-J., Tippetts, B. J. and Lillywhite, K. D. (2014). Date

maturity and quality evaluation using color distribution analysis and back
projection, Journal of Food Engineering 131: 161–169.

Zhang, L. and McCarthy, M. J. (2012). Measurement and evaluation of

tomato maturity using magnetic resonance imaging, Journal of Postharvest
Biology and Technology 67: 37–43.

Zhang, Y., Xie, X. and Cheng, T. (2010). Application of pso and svm in
image classification, 2010 3rd IEEE International Conference on Computer
Science and Information Technology (ICCSIT), Vol. 6, IEEE, pp. 629–631.

39