Computers in Industry 99 (2018) 69–82

Contents lists available at ScienceDirect

Computers in Industry
journal homepage: www.elsevier.com/locate/compind

A visible-range computer-vision system for automated, non-intrusive

assessment of the pH value in Thomson oranges
Sajad Sabzia , Juan Ignacio Arribasb,c,*
a
Department of Biosystems Engineering, College of Agriculture, University of Mohaghegh Ardabili, Ardabil, Iran
b
Department of Teoría de la Señal y Comunicaciones e Ingenieria Telematica, University of Valladolid, 47011 Valladolid, Spain
c
Castilla-León Neuroscience Institute (INCYL), University of Salamanca, 37007 Salamanca, Spain

A R T I C L E I N F O A B S T R A C T

Article history:
Received 12 August 2017 Fruit may be classified for the purposes of usage, packaging and marketing based on the pH (potential
Received in revised form 7 March 2018 of hydrogen) value – a numeric scale used to specify the acidity or basicity of an aqueous solution
Accepted 16 March 2018 measured in units of moles per liter of hydrogen ions. In this study, a new approach for the automated
Available online 25 April 2018 and non-intrusive estimation of the pH value of the Thomson navel orange (CRC 969, Citrus sinensis)
fruit is presented based on visible-range image processing, image feature extraction and with the use of
Keywords: hybrid imperialist competitive algorithm (ICA)-artificial neural network (ANN) regression. Image
Computer vision features studied include length, width, area, eccentricity, perimeter, blue-value, green-value, red-value,
Estimation
width, contrast, texture, roughness and several ratios thereof. Principal component analysis (PCA) is
Fruit grading
applied to reduce the number of dimensions without loss of important information and a cubic
Imperialist competitive algorithm
Neural network polynomial function of the mean square error (MSE) versus several factors is computed using the
Optimization response surface methodology (RSM) approach. Results for pH prediction are given and compared with
pH value true measured pH values over the entire 100 Thomson orange dataset, including estimated pH scatter
Regression regression plots and estimated pH boxplots. Cross validation is performed over 1000 repeated random
Response surface methodology trial experiments with uniform random train- and test-sample sets (80% training and 20% disjoint test
Thomson orange samples). In addition, we provide numerical results based on the levels achieved by response surface
methodology (RSM) evaluated over various error coefficients: the sum square error (SSE), the mean
absolute error (MAE), the coefficient of determination (R2), the root mean square error (RMSE), and MSE,
resulting in R2 = 0.843
0.043, MSE = 0.046
0.022, MAE = 0.166
0.039, SSE = 0.915
0.425, and
RMSE = 0.214
0.146, over the test set. The results demonstrate that such an automated pH-based
sorting system with machine vision using the hybrid ICA-ANN algorithm can accurately compute the pH
value of Thomson oranges without any contact with the fruit, and which has clear potential applications
in the food industry.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction
The estimation of internal quality of products has many
applications for example in the food industry to select suitable
classification, packing and marketing of fruit. Such determinations
currently use intrusive methods with corresponding impacts to the
products tested—non-intrusive and automated methods could
have significant impacts on productivity, quality and yield.
Machine (computer) vision can be applied together with pre-
processing algorithms to determine a set of indicators based on
* Corresponding author at: Department of Teoría de la Señal y Comunicaciones e
Ingenieria Telematica, University of Valladolid, 47011 Valladolid, Spain.
E-mail address: jarribas@tel.uva.es (J.I. Arribas).
visual properties that can predict specific quality features of the
product, to be fed in to a high-volume, non-intrusive sorting
system.
In the recent years, the interest, usage and importance of
machine learning in the food industry has grown rapidly and
several relevant papers have been published which compute
highly accurate and automatic classification of their respective
species under study; Zhao et al. [1] regarding damage recognition
on pest-infested oilseed rape leaves, [2] in fish species classifica-
tion by color, texture and multi-class support vector machine and
[3] in fruit classification by neural networks. Tests to measure
various physical qualities such as volume, weight, sphericity, pH,
taste and sugar content, can be done on different kinds of citrus
fruits such as orange [4]. Kondo et al. [5] demonstrated that the
quality of orange could be determined by a machine-vision and

https://doi.org/10.1016/j.compind.2018.03.016
0166-3615/© 2018 Elsevier B.V. All rights reserved.
70 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82
neural network system using different models to estimate at good
accuracy pH value and sugar content based on visual and physical
features.
An image segmentation method for apple sorting and grading
using support vector machines (SVM) and Otsu’s method was
investigated by [6]. This paper reports on the development of an
automatic adjustable algorithm for the segmentation of color
images, using linear SVM and Otsu’s thresholding method,
applied to apple sorting and grading. The method automatically
adjusts the classification hyperplane computed by using linear
SVM and needs minimum training and time. It also avoids the
problems caused by variations in the lighting conditions or the
color of the fruit. To evaluate the robustness and accuracy of the
proposed segmentation method, tests were conducted for 300
Red Delicious apples using three training input samples with
different color characteristics (orange, stripe, and dark red) and
their combination.
Iqbal et al. [7] studied predictions of moisture, color and pH
value in cooked, pre-sliced turkey hams by a near infrared (NIR)
hyperspectral imaging system in the NIR region ({900–1700 nm}
wavelength). S(PLS) regression, as a multivariate calibration
method to reduce the high dimensionality of the data and to
correlate the NIR reflectance spectra with quality attributes of the
samples under consideration. Instead of using a wide range of
spectra, the number of wavebands was reduced for more
stable, comprehensive and faster modelling in the subsequent
multispectral imaging system. With the help of PLS regression
analysis, nine different wavelengths (927, 944, 1004, 1058, 1108,
1212, 1259, 1362 and 1406 nm) were selected as the optimal
wavelengths for moisture prediction, eight different wavelengths
(927, 947, 1004, 1071, 1121, 1255, 1312 and 1641 nm) for pH
prediction and nine different wavelengths (914, 931, 991, 1115,
1164, 1218, 1282, 1362 and 1638 nm) were identified for color (a*)
prediction.
Vazquez-Cruz et al. [8] studied the application of neural networks
to estimate carotenoid content during ripening in tomato fruits
(Solanum lycopersicum). Six tomato ripening stages were consid-
ered. The relationship among color parameters (L*, a*, b*, and Hue),
maturity stages, and leaf area with the lycopene and b-carotene
concentration was analyzed with different regression models. Low
R2 values resulted, showing that lycopene and b-carotene content
were not well correlated with color during ripening stages. The
objective of this work was to provide a valid ANN model including
leaf area index and color readings as inputs to improve the
regression models for carotenoid estimations in tomatoes.
There are various image methods which can be potentially used
for these purposes; we have opted here for ordinary imaging
(visible range camera) due to its simplicity, easy availability of
camera equipment, its proven benefits, and appropriateness to be
used in an industrial environment. The use of non-destructive
methods to predict chemical features in fruits and crops is
welcome in agriculture and gardening as internal properties are
predicted with no damage done to the products. At the same time,
these methods increase the speed of grading and decrease the
amount of waste [9].
Several imaging systems include NIR, multispectral and
hyperspectral imaging, nuclear magnetic resonance imaging
Fig. 1. A system block diagram for the automatic non-intrusive computer vision sorting of Thomson orange based on pH value.
(NMR/MRI), X-ray computed tomography (mCT) [10] and visible
range imaging system [11], have been used successfully in non-
destructive fields. A visible imaging system is cheaper than other
systems and can be used easily. For instance, Mohapatra et al. [11]
presented a non-destructive method in banana grading based on
changes in the dielectric properties during ripening and image
processing. Bananas were stored in a maturity chamber at a
temperature of 25  C. During the ripening process some dielectric
properties including capacitance, relative permittivity and
impedance were measured. After measuring these properties,
one image was captured from each sample using a conventional
visible range camera. Three image processing methods were
used: Noise Reductant Local Binary Pattern (NRLBP), Local Binary
Pattern (LBP) and Completed Local Binary Pattern (CLBP), to
predict the degree of fruit maturity. As a final result, two
classifiers, Chi-Square distance/nearest-neighbor and Fuzzy-C
means (FCM), were successfully used for classification purposes.
Cárdenas–Pérez et al. [12], proposed an algorithm to predict
Young's modulus during the ripening stages of apple based on
color features, firmness and microstructure. Results showed that
Young's modulus during ripening stages decreased over time.
A significant correlation between Young’s modulus, firmness, skin
color difference, entropy and fractal dimensions was obtained
using Pearson's analysis. As a result, the determination coefficient
of Young's modulus predicted from firmness, skin color differ-
ence, entropy and fractal dimensions, was 0.95 using multiple
linear regression.
Rungpichayapichet et al. [13] studied the prediction of
physicochemical features of mango fruits including, firmness,
total soluble solids (TSS) and titratable acidity (TA) based on
hyperspectral images. Predicted models were developed based on
spectral data in visible and infrared ranges using partial least
squares regression. Results showed there exists a high correlation
between hyperspectral images and firmness, TSS and TA, so that
the values of the determination coefficient and root mean
squared error between hyperspectral images and firmness, TSS
and TA, were 0.81, 2.83 and 0.5 (determination coefficients), and
2, 0.81 and 0.24 (root mean square errors), respectively.
As noted above, researchers have focused in recent years on the
prediction of physicochemical properties based on color features,
and spectral data. Further development of these non-destructive
methods to non-intrusively evaluate physicochemical properties
in the different stages of fruit and crop ripening may allow
automated management, harvesting, handling, post harvesting
and warehousing to become a reality. Based on previous analyses,
we focused in this study on the most difficult and yet cheapest
non-destructive image prediction system, the visible range
imaging system (ordinary camera).
The specific aim of this study is to predict the pH value of
Thomson orange based on color and texture features from
ordinary images of fruit peel (exterior) using simple yet very
effective image processing techniques together with the ICA-ANN
methodology. Such a proposed methodology is novel and to our
best knowledge there are no similar reports published in the
literature so far.
Fig. 1 shows a system block diagram of the proposed system for
automatic and non-intrusive computer vision sorting of Thomson
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82
orange based on its pH value. The system comprises five main
stages: finding best lightning conditions, image acquisition, image
segmentation, feature extraction and pH value estimation.
2. Materials and methods
2.1. Samples
In this study, 100 samples of Thomson orange were purchased
from different shops on different days in order to obtain variability
(Fig. 2), and taken to the Physical Laboratory of the Faculty of
Agriculture Engineering, University of Razi in Kermanshah, Iran.
The pH value was measured with a pH meter (PH/ORP/Tempera-
ture meter, made in Taiwan). Total input data samples were divided
into two groups, including a training-set with 80 samples and a
test-set with 20 disjoint samples, performing cross validation
through 1000 repeated experiments with uniform random train-
and test-sets samples.
2.2. Image acquisition system
The machine vision system was composed of an image
acquisition system (Fig. 3), digital camera (BOSCH, Portugal), a
frame grabber (Pinnacle, China) and a personal computer (PC)
equipped with MatLab (ver. R2011a) and Microsoft Excel (ver.
2013) programs. The system provided images with a resolution of
352  288 pixels. The image acquisition system comprised three
types of lamps: LED, Fluorescent and Tungsten. A dimmer was built
for high and low light illumination conditions. The digital camera
was placed in the center of the image acquisition system. The
distance between sample and digital camera was fixed at 10 cm.
For image acquisition, only lamps of white LED with light intensity
of 50.33 lx were selected.
2.3. Image processing
2.3.1. Image processing stages
Fig. 4 shows a block diagram corresponding to image processing
operations carried out in this study. In turn, this block diagram
comprises two main parts: obtaining the best conditions of
photography and feature extraction.
2.3.2. Pre-processing and segmentation
Fig. 5 shows the histogram of a sample orange image in order to
determine the optimal threshold to segment samples from
background. There are 2 peaks in every subfigure in Fig. 5: first
peak on the left corresponds to fruit color pixels and second peak
corresponds to background pixels.
Fig. 2. A typical dataset sample Thomson orange image: (a) original image, (b) segmented orange image with background removal.
71
Fig. 3. Special purpose chamber used during image acquisition: lighting conditions are
controlled inside chamber, including LED, florescent and tungsten lamps light sources.
It can be inferred from Fig. 5 that for proper segmentation of
fruit from the background, a threshold that combines all RGB
components is necessary, as follows:
Rði; jÞ þ Gði; jÞ þ Bði; jÞ
< 30jRði; jÞ < 100 ð1Þ
4
Based on this threshold, if the sum of all components of one
pixel (after divide by 4) is less than 30 or the value of the red
component is less than 100, that pixel is considered to belong to
the background and omitted. Despite selection of the best
conditions for imaging, some noise still remains in the images;
thus, Canny and Laplacian filters were used for noise removal. The
Canny filter was used to detect the edges with local maximum
gradient f ðx; yÞ. The gradient is computed using a derivative of a
Gaussian filter. In this method, two thresholds are used to identify
strong and weak edges. In the Laplacian filter method, edges are
detected after filtering f ðx; yÞ using a Laplacian of Gaussian [14].
Fig. 6 shows the different stages of segmentation over four
different orange samples. From left to right, we show: original
image, segmentation with black background, binary image,
improved binary image, and segmented orange fruit.
2.3.3. Feature extraction
Based on an efficient algorithm that was designed and
implemented in the MatLab language, 14 features were extracted
from each orange sample. These features were: red component
value, green component value, blue component value, area,
72 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82

Fig. 4. The computer vision image analysis system block diagram.

2000
2000
1800

1600

1400 1500

1200

1000
1000
800

600
500
400

200

0 0

Fig. 5. Image segmentation typical histograms; removing background from fruit RGB image: (a) red component, (b) green component, (c) blue component.

perimeter, length, width, length/area, width/area, width/length,
eccentricity, the third moment, entropy, and homogeneity.
2.3.3.1. RGB color space. An RGB color image is a stack of three
single component color images. Indeed, each pixel of these images
has three elements, red, green and blue element channels. These
elements have limited values depending on the type of data class.
For example, if an RGB image is in the data class type double
(floating point double precision), uint8 (unsigned integer 8 bits) or
uint16 (unsigned integer 16 bits), it has the values between 0 and
1,0 and 255, 0 and 65535, respectively [14].
In general, other color spaces can be made from matrix
transformations from/to RGB color space [14]. We define the main
three components in RGB color space as follows:
- Red component value: this feature is defined by “the mean of the
first component computed over all pixels of each orange”.
- Green component value: this feature is defined by “the mean of
the second component computed over all pixels of each orange”.
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82 73

Fig. 6. Five different stages of image segmentation over four Thomson orange samples, (a)–(d), (top to bottom): (1) original image, (2) segmentation with black background,
(3) binary image, (4) improved binary image, and (5) final segmented orange image (left to right).

- Blue component value: this feature is defined by “the mean of the third moment was computed via Eq. (2), [14]:
the third component computed over all pixels of each orange”.
X
L1
m3 ¼ ðzi  mÞ3 pðzi Þ ð2Þ
i¼0
2.3.3.2. Area. Area is computed by counting the number of pixels
being zi a random variable of intensity values, pðzi Þ is the histogram
in a 2D region.
of intensity level values in one region, m is the mean intensity value
over that region of image and L the number of levels.
2.3.3.3. Perimeter. Perimeter is computed by counting the number
of pixels on the border of a closed region.
2.3.3.8. Entropy. Entropy is a measure of the random nature of
tissue. This feature is computed by:
2.3.3.4. Length. Length (pixels) is the major axis of the ellipse that
has the second moment equal to the area of a certain region. XL1
e¼ pðzi Þlog2 pðzi Þ ð3Þ
i¼0
2.3.3.5. Width. Width is the height (pixels) of minor axis of the
ellipse that has the second moment the same as the area of a
certain region. 2.3.3.9. Homogeneity. Homogeneity is a measure of image
uniformity. If all color levels are equal, homogeneity will be
2.3.3.6. Eccentricity. Eccentricity is defined as the ratio of the maximum. Homogeneity is computed by Eq. (4):
distance between the centers of the ellipse to its major axis length.
This value is between 0 and 1. An ellipse with eccentricity values of X
L1
U¼ p2 ðzi Þ ð4Þ
0 and 1, are a circle and a straight line, respectively. i¼0

2.3.3.7. The third moment. The third moment measures the

amount of symmetry in a histogram. The value of this moment
is zero for a symmetrical histogram, positive for a right-leaning
histogram and negative for a left-leaning histogram. In this study,
2.4. Principal component analysis (PCA)
As explained above, 14 features were extracted, so it is
appropriate and convenient to perform a dimensionality reduction
74 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82

to make a proper model for pH prediction. Principal components

analysis (PCA) is a statistical tool widely used for this purpose. In
particular, a reduction was made by PCA of the tuple of 14 values to
3 dimensions. In short, PCA is a way of identifying patterns in the
correlated data and expressing the data in such a way so as to
highlight their similarities and differences. The main advantage of
PCA is data compression by reducing the number of dimensions,
without loss of important information:

1. Data collection
2. Normalization of data
3. Calculation of covariance matrix
4. Interpretation of covariance matrix

Normalized data is then used to construct a variance-covariance

matrix N, as illustrated below:
2 3
C 1;1 C 1;2 : : : C 1;p
6 C 2;1 C 2;2 : : : C 2;p 7
6 7
6 : : : : : : 7
N¼6 6 :
7 ð5Þ
6 : : : : : 77
4 : : : : : : 5
C q;1 C q;2 : : : C q;p
 
where C q;p ¼ Cov X q ; X p , and X is set of input data. For instance:
Xn
ðX 1i  X 1 ÞðX 2i  X 2 Þ
CovðX 1 ; X 2 Þ ¼ i¼1
ð6Þ
ðn  1Þ

In the next step the eigenvectors and eigenvalues of the

covariance matrix are computed. In the final step, dimensionality
and thus feature vector size are reduced [15].

2.5. Imperialist competitive algorithm (ICA)

In this study, in order to train a Multilayer Perceptron (MLP)

ANN, an ICA algorithm was used. The pseudo code for the
proposed ICA method in the optimization of the aggregated
function can be written as shown in Fig. 7 flowchart, directly
adapted from [16].
Fig. 7. Flowchart of the ICA algorithm, adapted from [16].

2.6. Response surface methodology (RSM)

Calibration plays an important role in improving the perfor-
mance of the algorithm and, in some cases it is a necessary step
for properly developing the algorithms. RSM is a collection of
mathematical and statistical techniques developed by Box and
Wilson in the early 1950s. It is one of the best techniques used for
determining the parameter values of meta-heuristic algorithms
that are effective for the algorithm performance. If the variables
are assumed to be measurable, the response surface can be
expressed as y ¼ f ðx1 ; x2 ; . . . ; xk Þ; where k is the number of
variables. The goal here is to optimize the response variable y
under the premise that the independent variables are continuous
and controllable by experiments with negligible errors. Finding a
suitable approximation for a true functional relationship between
independent variables and the response surface is a prerequisite.
Neter [17] proposed a response surface function, y, with the
following polynomial form given in Eq. (7):
X
k XX X
k researchers.
y ¼ b0 þ bj xj þ bij xi xj þ bjj x2j þ e
j¼1 i<j j¼1
where, y is the predicted response, b0 is a constant, bj is the linear
coefficient, bij is the cross-product coefficient, and bjj is the
squared coefficient [18]. In this study, five parameters include: the
Number of Countries (A), Iteration (B), Revolution Rate (C), Uniting
Threshold (D), and Assimilation Coefficient (E), optimally tuned by
the RSM algorithm.
3. Results and discussion
3.1. Data generation and settings
An experiment was conducted to test the performance of the
hybrid ICA-ANN using RSM. This method determines the optimal
number of experiments by computing the number of factors and
their levels in order to set optimal factors of the ICA-ANN approach.
Table 1 shows the number of factors and their levels. As you
can see, there are five factors and two levels for this experiment.
The selected factors are: Number of Countries (A), Iteration (B),
Revolution Rate (C), Uniting Threshold (D) and Assimilation
Coefficient (E). These factors and their levels were selected
based on past experience and previous results from other
ð7Þ
3.2. Hybrid ICA-ANN algorithm parameters tuning
It is known that a good choice of parameters has a significant
impact on performance of algorithms and that suitable parameter
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82 75

Table 1
RSM methodology applied to the ICA-ANN approach: factors and their optimal levels. RSM method is used for estimating the optimal number of experiments by computing
the number of factors and their levels in order to set optimal factors of the ICA-ANN approach. Table shows the number of factors (and their levels), comprising: Number of
Countries (A), Iteration (B), Revolution Rate (C), Uniting Threshold (D) and Assimilation Coefficient (E).

Level Number of Countries (A) Iteration (B) Revolution Rate (C) Uniting Threshold (D) Assimilation Coefficient (E)
2 375 500 0.35 0.030 2.5
Lower (1) 300 400 0.30 0.025 2.0
0 225 300 0.25 0.020 1.5
Upper (1) 150 200 0.20 0.015 1.0
2 75 100 0.15 0.010 0.5

values depend on the type of problem. Most research has been
conducted using evolutionary algorithms with fixed parameter
values after some preliminary initialization experiment or values
have been fixed with reference to the values of similar previous
work. Calibration is important in the performance improvement of
algorithms and, in some cases, it is a necessary step in the
development of algorithms. For the purpose of calibrating the
algorithms, several methods have been used in the literature [19].
In this section, the behavior of different parameters of the
proposed hybrid ICA-ANN algorithm is studied based on RSM –
values and experimental results, namely mean square errors
(MSE), are presented in Table 2. There are a total of 32 experiments
considered in this table, since RSM proposed 32 experiments to set
parameters based on the number of factors under study. There are
three different point types in Table 2: factorial, central and axial
points. As one can see, if all parameters are set equal to 0 level,
point type will be central, if parameters are equal to 1 or 1 level,
then point type will be factorial, otherwise point type will be axial.
The result of each experiment was evaluated with the use of MSE
criterion. In order to study the effectiveness and efficiency of the
algorithms, a comparative performance measure called the relative
percentage deviation (RPD), is used. This metric is computed using
Eq. (8):
Algorithmsol  Minsol
RPD ¼  100 ð8Þ
Minsol
where Algorithmsol is the result obtained for a given algorithm and
Minsol is minimum result obtained among all the experiments [20].
At this point, results obtained in the last column in Table 2
(RPD) were computed using the Design-Expert 7.00 software to
find the best model. This software proposed the cubic model as
optimal. As is well-known, a cubic model has a maximum of three
free parameters. Estimated regression coefficients of the cubic
model are given in Table 3. This software, among all 56 possible
regressions, selected the 25 best regressions. Table 3 shows the
impact each potential regression has on the final model. The
estimated regression models can be stated as follows:

Table 2
Experimental results of central composite design for the numerical examples and MSE. RSM results of experiments, including MSE error. There are a total of 32 experiments
considered in this table. There are three different point types: factorial, central and axial points. The result of each experiment was evaluated with MSE criterion. In order to
study the effectiveness and efficiency of the algorithms, a comparative performance measure called RPD is used and computed using (8) with the help of the Design-Expert
7.00 software, so as to find the best optimal model, resulting in the cubic model.

Factors Point type A B C D E MSE RPD

Experiments
1 Fact 1.00 1.00 1.00 1.00 1.00 0.220424 66.79581088
2 Fact 1.00 1.00 1.00 1.00 1.00 0.229904 73.96936861
3 Fact 1.00 1.00 1.00 1.00 1.00 0.387152 192.9596223
4 Axial 2.00 0.00 0.00 0.00 0.00 0.197372 49.35226103
5 Center 0.00 0.00 0.00 0.00 0.00 0.240475 81.96849083
6 Axial 0.00 0.00 0.00 0.00 2.00 0.132152 0
7 Center 0.00 0.00 0.00 0.00 0.00 0.269653 104.0476118
8 Axial 0.00 2.00 0.00 0.00 0.00 0.278034 110.3895514
9 Center 0.00 0.00 0.00 0.00 0.00 0.230918 74.73666687
10 Fact 1.00 1.00 1.00 1.00 1.00 0.170639 29.12328228
11 Fact 1.00 1.00 1.00 1.00 1.00 0.139724 5.729765724
12 Axial 0.00 0.00 0.00 2.00 0.00 0.257156 94.59107694
13 Fact 1.00 1.00 1.00 1.00 1.00 0.319787 141.9842303
14 Fact 1.00 1.00 1.00 1.00 1.00 0.296879 124.6496459
15 Fact 1.00 1.00 1.00 1.00 1.00 0.238934 80.80240935
16 Fact 1.00 1.00 1.00 1.00 1.00 0.188726 42.80979478
17 Center 0.00 0.00 0.00 0.00 0.00 0.191985 45.27589442
18 Axial 0.00 0.00 2.00 0.00 0.00 0.216183 63.58662752
19 Axial 2.00 0.00 0.00 0.00 0.00 0.243553 84.29762697
20 Fact 1.00 1.00 1.00 1.00 1.00 0.300538 127.4184273
21 Center 0.00 0.00 0.00 0.00 0.00 0.369456 179.5689812
22 Fact 1.00 1.00 1.00 1.00 1.00 0.151207 14.41900236
23 Fact 1.00 1.00 1.00 1.00 1.00 0.135624 2.62727768
24 Fact 1.00 1.00 1.00 1.00 1.00 0.173509 31.29502391
25 Axial 0.00 0.00 0.00 2.00 0.00 0.280278 112.0875961
26 Axial 0.00 2.00 0.00 0.00 0.00 0.293017 121.7272535
27 Axial 0.00 0.00 0.00 0.00 2.00 0.195875 48.21947454
28 Fact 1.00 1.00 1.00 1.00 1.00 0.341757 158.609026
29 Center 0.00 0.00 0.00 0.00 0.00 0.208988 58.14213935
30 Fact 1.00 1.00 1.00 1.00 1.00 0.140287 6.155790302
31 Fact 1.00 1.00 1.00 1.00 1.00 0.177391 34.2325504
32 Axial 0.00 0.00 2.00 0.00 0.00 0.206257 56.07557964
76 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82

Table 3
Estimated regression coefficients. Estimated regression coefficients of the cubic optimal model. Among all 56 possible regressions, the 25 best regressions were selected and
shown in this table. Table also shows the impact each potential regression has in final model. The estimated regression models can be stated as given by Eq. (9).

Term Coefficients Standard Error F-value p-value

A–A Ncountries 0.11 2.278 2.18  105 0.9964

B–B Iteration 18.19 2.278 0.64 0.4551
C–C Revolution Rate 3.84 2.278 0.028 0.8716
D–D Uniting Threshold 2.57 2.278 0.013 0.9137
E–E Assimilation Coefficient 16.70 2.278 0.54 0.4911
AB Ncountries  Iteration 1.47 1.611 8.31 103 0.9303
AC Ncountries  Revolution Rate 2.90 1.611 0.032 0.8630
AD Ncountries  Uniting Threshold 2.63 1.611 0.027 0.8755
AE Ncountries  Assimilation Coefficient 6.69 1.611 0.17 0.6922
BC Iteration  Revolution Rate 10.84 1.611 0.45 0.5259
BD Iteration  Uniting Threshold 12.08 1.611 0.56 0.4816
BE Iteration  Assimilation Coefficient 13.74 1.611 0.73 0.4263
CD Revolution Rate  Iteration 23.30 1.611 2.09 0.1981
CE Revolution Rate  Assimilation Coefficient 27.66 1.611 2.95 0.1368
DE Uniting Threshold  Assimilation Coefficient 1.84 1.611 0.013 0.9128
A2 Number of Countries  Number of Countries 18.38 1.190 2.39 0.1732
B2 Iteration  Iteration 2.95 1.190 0.062 0.8123
C2 Revolution Rate  Revolution Rate 5.89 1.190 0.25 0.6379
D2 Uniting Threshold  Uniting Threshold 2.49 1.190 0.044 0.8410
E2 Assimilation Coefficient  Assimilation Coefficient 0.20 1.190 2.8  104 0.9872
A2 B Number of Countries  Number of CountriesIteration 38.17 2.790 1.87 0.2203
A2 C Number of Countries  Number of CountriesRevolution Rate 1.23 2.790 1.13  103 0.9663
A2 D Number of Countries  Number of Countries  Uniting Threshold 7.45 2.790 0.071 0.7984
A2 E Number of Countries  Number of Countries  Assimilation Coefficient 13.56 2.790 0.24 0.6442
AB2 Ncountries  Iteration  Iteration 16.40 2.790 0.35 0.5781

R1 ¼ 94:25 þ 0:11A þ 18:19B  3:84C þ 2:57D þ 16:7E  1:47AB

3.3. Prediction of pH values
þ2:9AC  2:63AD þ 6:69AE þ 10:84BC þ 12:08BD
þ13:74BE  23:3CD  27:66CE We propose to use back-propagation and MLP architecture for
the ANN to estimate pH. ICA was used to optimize the MLP
þ1:84DE  18:38A2 2:95B2 5:89C2 architecture determining the optimal number of neurons, the
number of hidden layers (model selection), the nonlinear transfer
þ2:49D2 0:2E2 38:17A2 B
function, the back-propagation training function and the weights/
þ1:23A2 C  7:45A2 D  13:56A2 E  16:4AB2 ð9Þ bias learning function. The ICA optimized parameters for the
proposed ANN for pH estimation are presented in Table 4. During
At the end, RSM proposed that the optimal values of A, B, C, D
computer simulations, in order to compute statistically significant
and E factors were 150, 185, 0.29, 0.018 and 1.8 respectively. We
results, we divided input orange data into training (80 samples)
chose ICA-ANN methodology to run with these estimated optimal
and test (20 samples) sets uniformly random, and repeated
factor values to obtain the best results possible. The dependence
simulations 1000 times, every time with different uniform random
between Ncountries-Iteration and dependence between Revolution
training and test sets so as to achieve proper cross validation.
Rate-Uniting Threshold, are shown in Fig. 8. Both Ncountries-
Thus, since we have 100 different input orange pH samples,
Iteration and Revolution Rate-Uniting Threshold dependences
each orange was used on average 200 times in the test set. We
comprise 4 factors corresponding to ICA-ANN that were tuned
now compute a regression scatter plot for actual and estimated
using RSM methodology. As it can be seen in Fig. 8(a), increasing
pH values, compute mean and standard deviation values for both
Ncountries and Iteration results in decreasing MSE. On the other
the predicted pH values (as compared to actual measured true pH
hand, Fig. 8(b) shows that decreasing Revolution Rate and Uniting
values) and for each error coefficient. In addition, we also
Threshold, MSE is also decreased.
compute the actual minus estimated pH boxplots for each input

Table 4
ICA optimal MLP ANN parameters used in Thomson orange non-intrusive pH estimation.

number of neurons first layer:5; second layer:4

number of layers 2
transfer function first layer: tansig; second layer: tansig
backpropagation network training function trainlm
backpropagation weight/bias learning function learnh

Table 5
Regression and error coefficients for the automatic Thomson orange pH estimation model. Mean and standard deviation of regression and error coefficient for the 1000
uniform random test set samples simulations averaged. On average, each orange sample is used 200 times.

R2 SSE MAE MSE RMSE

mean (best) 0.843 (0.962) 0.915 (0.533) 0.166 (0.126) 0.046 (0.027) 0.214 (0.163)
standard deviation 0.043 0.425 0.039 0.022 0.146
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82 77

Fig. 8. R1 MSE error surface computed dependence between various critical optimization algorithm parameters: (a) Iteration and Ncountries, (b) Uniting Threshold and
Revolution Rate.

Estimated pH
Measured pH

Fig. 9. A visual example for comparison purposes, comprising 20 measured (true) and estimated (predicted) Thomson orange pH values, over the test set.
78 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82

Table 6
True and mean predicted pH values for the 100 Thomson oranges, over the test set. Standard deviation of estimated pH values is also included, for the 1000 uniform random
test set samples simulations averaged: on average, each orange # is used 200 times inside the test set.

orange # true pH predicted pH orange # true pH predicted pH orange # true pH predicted pH

1 3.95 4.08
0.18 34 5.01 4.9
0.22 68 4.77 4.73
0.23
35 5.54 5.35
0.27 69 5.33 5.33
0.27
2 4.26 4.16
0.19 36 3.79 4.08
0.17 70 4.93 4.90
0.26
3 4.24 4.21
0.23 37 5.61 5.21
0.32 71 5.00 4.99
0.25
4 3.75 4.48
0.27 38 5.17 5.00
0.39 72 3.83 4.08
0.20
5 4.46 4.15
0.20 39 4.49 4.38
0.25 73 5.18 5.25
0.25
6 4.14 4.14
0.21 40 3.69 3.98
0.21 74 4.62 4.42
0.20
7 3.96 4.39
0.26 41 4.98 4.84
0.23 75 5.17 5.07
0.23
8 4.43 4.43
0.25 42 4.74 4.63
0.35 76 4.05 4.17
0.19
9 4.42 4.02
0.18 43 4.19 4.21
0.19 77 4.46 4.31
0.22
10 3.72 4.05
0.19 44 4.06 4.16
0.22 78 3.61 4.01
0.18
11 5.65 5.43
0.36 45 3.88 4.08
0.17 79 5.01 4.76
0.23
12 4.37 4.18
0.20 46 4.33 4.26
0.22 80 5.71 5.52
0.26
13 4.32 4.21
0.18 47 4.12 4.14
0.20 81 4.33 4.44
0.19
14 5.21 4.97
0.23 48 4.41 4.27
0.19 82 3.66 3.96
0.16
15 5.02 5.13
0.24 49 4.24 4.18
0.21 83 4.21 4.26
0.19
16 3.84 4.03
0.19 50 4.96 4.88
0.23 84 3.61 3.95
0.17
17 3.98 4.10
0.17 51 3.89 4.11
0.18 85 4.23 4.25
0.18
18 4.06 4.19
0.20 52 4.05 4.21
0.19 86 4.10 4.32
0.22
19 3.48 3.88
0.19 53 5.00 5.14
0.23 87 4.45 4.46
0.20
20 4.48 4.30
0.18 54 5.21 5.37
0.31 88 6.31 5.79
0.24
21 5.27 4.89
0.25 55 4.54 4.46
0.20 89 4.50 4.61
0.24
22 5.35 5.23
0.24 56 4.79 4.62
0.30 90 5.21 5.30
0.25
23 4.24 4.25
0.18 57 4.72 4.74
0.25 91 5.01 5.05
0.24
24 4.96 4.80
0.23 58 4.86 4.87
0.23 92 5.93 5.56
0.23
25 6.26 5.70
0.22 59 4.53 4.35
0.25 93 5.16 5.09
0.23
26 4.91 4.90
0.23 60 4.8 4.73
0.23 94 4.20 4.44
0.22
27 4.73 4.57
0.22 61 4.69 4.66
0.23 95 4.57 4.41
0.20
28 5.41 5.12
0.36 62 5.54 5.35
0.25 96 5.40 5.31
0.24
29 3.95 4.01
0.20 63 5.58 5.14
0.24 97 5.33 5.18
0.24
30 5.76 5.32
0.29 64 6.23 5.68
0.23 98 5.68 5.46
0.25
31 5.01 4.95
0.34 65 5.00 4.78
0.22 99 4.78 4.71
0.23
32 4.14 4.15
0.18 66 3.71 4.04
0.19 100 5.73 5.49
0.25
33 4.00 4.09
0.19 67 5.72 5.61
0.30

Data
Fit
X=Y
6

5.5

4.5

3.5
3.5 4 4.5 5 5.5 6

Fig. 10. Scatter plot regression analysis between mean estimated and actual (measured) Thomson orange pH values (test set). The linear best fit is also depicted in blue color
line. 1000 uniform test sets where used to compute mean estimated pH values for each of the 100 oranges (depicted as a circular point in figure). On average, each orange is
used 1000 repetitions *20 samples in test set/100 input oranges = 200 times, inside the test set. Note: for comparison purposes, the X = Y straight line is also depicted as a
dotted line in figure. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82 79

Fig. 11. pH boxplots for the 100 Thomson orange database. Actual-estimated pH values (test set): 1000 uniform random test set samples simulations averaged; on average, each input orange # is used 1000*20/100 = 200 times.
orange sample as well as boxplots for each error coefficient as
defined in the next paragraph.
We define five typical error coefficient measures to be
employed as performance indicators to evaluate prediction
capability of the proposed model [21,22]: the Sum square error
(SSE), the mean absolute error (MAE), the coefficient of
determination (R2), the root mean square error (RMSE), and
the mean squared error (MSE):
8 Xn 9
< ðXk  X0 Þ2 =
2
R ¼ 1  Xn k¼1
ð10Þ
: ðXk  Xm Þ2 ;
k¼1

1X n
Xm ¼ X ð11Þ
n k¼1 k

1X n
MSE ¼ ðX  X0 Þ2 ð12Þ
n k¼1 k

X
n
SSE ¼ ðXk  X0 Þ2 ð13Þ
k¼1

1X n
MAE ¼ jX  X0 j ð14Þ
n k¼1 k
s

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n
u 1X
RMSE ¼ t ðX  X0 Þ2 ð15Þ
n k¼1 k

Where Xk are the actual values, Xo are the forecast values, X0 are
the average of experimental values, Xm is the mean of the actual
true values, and n is the number of samples.
Table 5 shows orange pH estimation regression errors based
on the levels achieved by RSM. The results shown indicate that
our hybrid model yielded a reliable performance including
proper generalization and prediction capacities. Fig. 9 and
Table 6 show pH prediction results for our data making use of
the hybrid ICA-ANN algorithm and using 1000 uniform random
test sets. The visual example given in Fig. 9, shows that the
proposed hybrid ICA-ANN model had good learning capability
for the input training samples (80 samples) and simultaneously
excellent generalization performance over the disjoint test
samples (20 samples). ICA was a good compromise for
guaranteeing improvement in terms of both stability and
accuracy and was a suitable and effective method for predicting
the pH value in Thomson orange. Regression analysis was
conducted on the estimated pH values with respect to the
measured pH obtained from the entire experiment sets. Fig. 10
depicts results of regression analysis, resulting in a linear
regression coefficient between measured and estimated pH
values for Thomson orange of R = 0.96, which for limited
resolution input images (as could be found in a real-time
industrial environment) and a non-intrusive pH estimation
approach, we consider a rather remarkable quality of fit.
Nevertheless, it can be seen that the system tends to slightly
truncate the value of the estimated pH for extreme values in both
maxima (reducing estimated pH values of those oranges close to a
6.5 pH value) and minima (increasing estimated pH values of those
oranges close to a 3.5 pH value).
In Fig. 11 we have computed for each of the 100 input oranges
the actual minus the estimated pH value boxplots (test set),
80 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82

0.9

0.8

0.7

0.6

0.5

0.4

SSE R R^2

(a)
0.45

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
MSE MAE RMSE

Fig. 12. Boxplots for the various error coefficients used, for 1000 uniform random test set samples simulations averaged; thus, on average, each orange # is used 200 times
inside the test set: (a) SSE (b) R and R2, and (c) MSE, MAE and RMSE.

using 1000 uniform random test set samples simulations results. As can be seen from Fig. 11, the accuracy of the non-
averaged; Thus, on average, each input orange sample is used intrusive system is again remarkable, showing small mean
200 times in the test sets, so as to achieve statistically valid dispersion (variance) errors for the whole dataset and with

Fig. 13. Layout of a potential automatic and non-destructive Thomson orange pH sorting system prototype: belt moves from right to left in image while oranges are classified
in either, low, medium or high pH value ranges.
 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82
limited bias; it should be noted how in subtraction between the
actual and estimated pH value boxplots exhibit a median value
close to zero, what is nice.
Finally, we show in Fig. 12 various boxplots for several error
and accuracy coefficients, as defined above, including boxplots
for the MSE, MAE, MRSE, SSE, R and R2, again for the same 1000
uniform random test set experiments introduced earlier. One
can see how the shape of those coefficient boxplots is rather
remarkable implying the accuracy with which the non-intrusive
automatic computer vision system is predicting orange pH
values.
Regarding computational efficiency, on average each image is
processed in 0.53 s on an off-the-shelf PC with an Intel Core i3 CFI
processor, 330 M at 2.13 GHz, 4 GB of RAM and Windows 10.
Image segmentation takes 0.20s, feature extraction 0.15s, and
application of the ANN 0.15s, on average for each orange input
sample image.
3.4. pH grading system proposal prototype
A Thomson orange pH value grading system prototype is
depicted in Fig. 13. Oranges with different pH values are placed on
conveyor belt (moving right to left in figure) and the camera takes
images from them. The captured images are then transferred to
the computer for analysis by our proposed computer vision
algorithm developed in MatLab. After the analysis, the fruits
are divided into three groups: fruit with high, medium and low
pH values. Three different lanes are configured to move oranges
with different pH grades to the corresponding processing or
packaging area.
Fig. 14 shows final computer vision system setup used for
proper automatic estimation of Thomson orange pH values.
4. Conclusions
A novel method was proposed for automatic and non-intrusive
estimation of pH value in Thomson oranges based on computer
vision in the visible range. By automatic we mean that the value of
Fig. 14. Final computer vision system setup overview to automatically estimate Thomson orange pH value.
81
pH for each orange is estimated through regression from the image
data without human intervention.
The novelty of the approach lies in the prediction of pH of
Thomson oranges based on color and texture features that may be
one of the most difficult and at the same time simplest non-
destructive prediction system approaches in the visible range
(available with an ordinary camera). The time used in non-intrusive
pH value prediction was 0.53 s of which image segmentation took
0.20 s, feature extraction 0.15 s, and application of the ICA-ANN
0.15 s, on average for each orange image input sample.
To conclude, we recap that the method described in this paper
includes image pre-processing to compute 14 relevant attributes
from the orange population to estimate pH value by the hybrid
ICA-ANN algorithm. Specifically, using PCA and RSM methods the
algorithm predicted orange pH value with good accuracy in a non-
intrusive fashion. As proven by results in computer simulations
provided above, accuracy in pH value estimation is rather high,
particularly considering the type of low-resolution visible range
input images. Accurate classification in low-mid-high pH values
of Thomson oranges could be valid for potential industrial real-
time applications and/or non-intrusive automatic fruit pH sorting
systems.
We believe accuracy in the performance was possible due to the
following key factors:
1. The hybrid ICA-ANN algorithm, like most other optimization
algorithms, is mainly influenced by values of its parameters.
These parameters can be set manually or by using different
optimisation approaches. By using the RSM approach, a large
number of critical variables are tuned through a small number of
experiments.
2. The automatic and non-intrusive Thomson orange pH sorting
system and prototype proposed in this study could be
potentially used in industry, but in order to achieve a lower
error sorting system, the following items should be taken into
consideration:
a. Use of a camera with a higher resolution.
82 S. Sabzi, J.I. Arribas / Computers in Industry 99 (2018) 69–82
b. Placing the camera at a suitable fixed distance to fruit in such a
way that the real size of fruit is always shown.
c. Use of appropriate lighting conditions without any disturbances
or changes during the whole images acquisition procedure.
d. Choosing a suitable color for the background avoiding any light
reflections.
Acknowledgments
J. I. Arribas is very grateful to physicist Dr. Luis del Pozo,
Geneva, Switzerland, for reviewing and commenting on the
manuscript. This study was funded in part by Ministerio de
Economía y Competitividad, Spain, under grant number
TEC2013-44194-P.
References
[1] Y. Zhao, Y. He, X. Xu, A novel algorithm for damage recognition on pest-infested
oilseed rape leaves, Comput. Electron. Agric. 89 (2012) 41–50.
[2] J. Hu, D. Li, Q. Duan, Y. Han, G. Chen, X. Si, Fish species classification by color,
texture and multi-class support vector machine using computer vision,
Comput. Electron. Agric. 88 (2012) 133–140.
[3] Y. Zhang, P. Phillips, S. Wang, G. Ji, J. Yang, J. Wu, Fruit classification by
biogeography-based optimization and feedforward neural network, Expert
Syst. 33 (3) (2016) 239–253.
[4] A. Adelkhani, B. Beheshti, S. Minaei, P. Javadikia, M. Ghasemi-Varnamkhasti,
Taste characterization of orange using image processing combined with ANFIS,
Measurement 46 (2013) 3573–3580.
[5] N. Kondo, U. Ahmad, M. Monta, H. Murase, Machine vision based quality
evaluation of Iyokan orange fruit using neural networks, Comput. Electron.
Agric. 29 (2000) 135–147.
[6] A. Mizushima, R. Lu, An image segmentation method for apple sorting and
grading using support vector machine and Otsu’s method, Comput. Electron.
Agric. 94 (2013) 29–37.
[7] A. Iqbal, D.W. Sun, P. Allen, Prediction of moisture, color and pH in cooked, pre-
sliced turkey hams by NIR hyperspectral imaging system, J. Food Eng. 117
(2013) 42–51.
[8] M.A. Vazquez-Cruz, S.N. Jimenez-Garcia, R. Luna-Rubio, L.M. Contreras-
Medina, E. Vazquez-Barrios, E. Mercado-Silva, I. Torres-Pacheco, R.G. Guevara-
Gonzalez, Application of neural networks to estimate carotenoid content
duringripening in tomato fruits (Solanum lycopersicum), Scientia
Horticulturae 162 (2013) 165–171.
[9] P.J. Ramos, F.A. Prieto, E.C. Montoya, C.E. Oliveros, Automatic fruit count on
coffee branches using computer vision, Comput. Electron. Agric. 137 (2017) 9–
22.
[10] E. Arendse, O.A. Fawole, L.S. Magwaza, U.L. Opara, Non–destructive prediction
of internal and external quality attributes of fruit with thick rind: a review, J.
Food Eng. (2017) 1–13 article in press.
[11] A. Mohapatra, S. Shanmugasundaram, R. Malmathanraj, Grading of ripening
stages of red banana using dielectric properties changes and image processing
approach, Comput. Electron. Agric. 143 (2017) 100–110.
[12] S. Cárdenas–Pérez, J.V. Méndez–Méndez, J.J. Chanona–Pérez, A. Zdunek, N.
Güemes–Vera, G. Calderón–Domínguez, F. Rodríguez–González, Prediction of
the nanomechanical properties of apple tissue during its ripening process
from its firmness, color and microstructural parameters, Innov. Food Sci.
Emerg. Technol. 39 (2017) 79–87.
[13] P. Rungpichayapichet, M. Nagle, P. Yuwanbun, P. Khuwijitjaru, B. Mahayothee,
J. Muller, Prediction mapping of physicochemical properties in mango by
hyperspectral imaging, Biosyst. Eng. 159 (2017) 109–120.
[14] R.C. Gonzalez, R.E. Woods, S.L. Eddins, Digital Image Processing Using MATLAB,
Prentice Hall, 2004.
[15] S. Datta, G. Nandi, A. Bandyopadhyay, P.K. Pal, Application of PCA–based hybrid
Taguchi method for correlated multicriteria optimization of submerged arc
weld: a case study, Int. J. Adv. Manuf. Technol. 45 (2009) 276–286.
[16] E. Atashpaz–Gargari, C. Lucas, Imperialist competitive algorithm: an algorithm
for optimization inspired by imperialistic competition, IEEE Congress on
Evolutionary Computation, (2007) .
[17] J. Neter, Applied linear statistical models, WCB McGraw-Hill, New York,
1996.
[18] M. Rabiee, M. Zandieh, A. Jafarian, Scheduling of a no-wait two-machine flow
shop with sequence-dependent setup times and probable rework using robust
meta-heuristics, Int. J. Prod. Res. (2012) 1–19.
[19] F. Jolai, M. Rabiee, H. Asefi, A novel hybrid meta–heuristic algorithm for a no-
wait flexible flow shop scheduling problem with sequence dependent setup
times, Int. J. Prod. Res. (2012) 1–20.
[20] E. Shokrollahpour, M. Zandieh, B. Dorri, A novel imperialist competitive
algorithm for bi-criteria scheduling of the assembly flowshop problem, Int. J.
Prod. Res. 49 (2011) 3087–3103.
[21] S. Liu, L. Xu, D. Li, Q. Li, Y. Jiang, H. Tai, L. Zeng, Prediction of dissolved oxygen
content in river crab culture based on least squares support vector regression
optimized by improved particle swarm optimization, Comput. Electron. Agric.
95 (2013) 82–91.
[22] S. Sabzi, P. Javadikia, H. Rabani, A. Adelkhani, Mass modeling of bam orange
with ANFIS and SPSS methods for using in machine vision, Measurement 46
(2013) 3333–3341.
Sajad Sabzi received his PhD degree in Biosystem
Engineering from University of Mohaghegh Ardabili, Iran
in 2017. He is now working as a Postdoctoral researcher in
University of Mohaghegh Ardabili, Iran. His research
interests include Precision agriculture, Image processing,
Artificial Intelligence and Machine Vision.
Juan Ignacio Arribas (PhD) received the MSc and PhD
degrees both in Electrical Engineering from Universidad
Valladolid, Spain, in 1996 and 2001, respectively. In 1997
he was an FPI Postgraduate Research Fellow at the
Department of Science, Spain. Since 1999, he has been
with the Department of Electrical Engineering, Universi-
dad Valladolid, Spain, currently as an Associate Professor.
During 1998 and 2009, he was a Visiting Research
Associate at the University of Maryland, Baltimore,
Maryland. In 2010, he visited the Barrrow Ne urological
Institute, Phoenix, Arizona. His research interests are
Machine Learning, Pattern Recognition, Neural Networks,
Expert Systems and their applications to Computer Aided
Diagnosis and Computer Vision. Dr. Arribas is Member Castilla – Leon Institute of
Neuroscience (INCYL), University of Salamanca, Spain, Editorial Board Member
Scientific Data (Nature Publishing Group), Editorial Board Member Journal
Renewable Energy Research (IJRER), and Associate Editor Expert Systems (Wiley).