Classification of Trifolium Seeds by Computer Vision Methods
RECEP ERYIGIT, YILMAZ AR, BULENT TUGRUL∗
Department of Computer Engineering
Ankara University
Golbasi, Ankara, TURKEY
Abstract: - Traditional machine learning methods have been extensively used in computer vision applications.
However, recent improvements in computer technology have changed this trend. The dominance of deep learn-
ing methods in the field is observed when state-of-the-art studies are examined. This study employs traditional
computer vision methods and deep learning to classify five different types of Trifolium seeds. Trifolium, the lead-
ing food for nutritious dairy products, plays an essential role in livestock in some parts of the world. First, an
image data set consisting of 1903 images belonging to five different species of Trifolium was created. Descriptive
and quantitative morphological features of each species are extracted using image-processing techniques. Then a
feature matrix was created using eight different features. After feature selection and transformation, unnecessary
and irrelevant features were removed from the data set to build more accurate and robust classification models.
Four common and frequently applied classification algorithms created a prediction model in the seed data set. In
addition, the same dataset was trained using VGG19, a convolutional neural network. Finally, the performance
metrics of each classifier were computed and evaluated. The decision tree has the worst accuracy among the four
traditional methods, 92.07%. On the other hand, Artificial Neural Network has the highest accuracy with 94.59%.
As expected, VGG19 outperforms all traditional methods with 96.29% accuracy. However, as the results show,
traditional methods can also produce results close to the deep learning methods.
Key-Words: - Classification, Computer Vision, Machine Learning, Deep Learning, Trifolium
5HFHLYHG-XO\5HYLVHG)HEUXDU\$FFHSWHG0DUFK3XEOLVKHG0DUFK
1 Introduction
Computer vision is extracting valuable information
from digital images or videos acquired from an elec-
tronic camera to decide and perform an action based
on that information. Due to the latest developments, it
has been widely used in various applications, such as
autonomous vehicle systems [1, 2], health care treat-
ments [3, 4] and quality control processes on produc-
tion lines [5, 6]. A traditional computer vision ap-
plication requires the following procedures: image
acquisition, prepossessing, segmentation, feature ex-
traction, recognition and knowledge base. In the fea-
ture extraction stage, some high-dimensional quanti-
tative data of the image is collected for use in the fol-
lowing steps. These features are edges, corners, blobs
and ridges. They should not be sensitive to rotation,
scale, translation, and contrast changes. The selected
features must contain relevant information so that the
requested task succeeds. For example, a face recog-
nition system measures the distinctive organs of a hu-
man face, such as the nose, eyes and eyebrows [7].
When applying traditional computer vision meth-
ods to a problem, much of the time is spent preparing
the data and extracting feature sets that require ex-
pert knowledge of the problem domain. In addition,
these human-crafted features can be incomplete and
specific. The cycle is repeated until the desired ac-
curacy and consistency are achieved. However, deep
learning has wholly altered the time distribution of the
learning process. Deep learning applications spend
most of their time modelling data with neural net-
work architectures that automatically convert raw in-
puts into valuable features.
Precision farming can be defined as applying the
latest technology and science in agriculture. Tech-
nologies such as Machine Learning, Cloud Comput-
ing, Big Data, and the Internet of Things (IoT) help
farmers make more accurate and conscious decisions
about the consequences of their actions [8,9]. Thanks
to the sensors placed in the fields, critical choices such
as irrigation time and amount, the need for pesticide
use and when to start harvesting can be automated by
eliminating the human factor. Moreover, data such as
the type, quality and quantity of crops harvested in the
coming season can be obtained using digital images
from satellites. Thus, administrations can take nec-
essary precautions against problems in the food mar-
ket [10]. In addition, studies based on image process-
ing and computer vision methods try to detect crop
diseases as early as possible to improve product effi-
ciency.
The genus Trifolium includes more than 200
species, some of which are rich in calcium, phospho-
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
313
Volume 22, 2023
rus and various vitamins, so they are important in the
livestock economy [11–13]. A study conducted in
1976 revealed that Trifolium species are among the
most valuable forage legumes in all parts of the world
except the tropics [14]. Trifolium species are impor-
tant nitrogen fixers in temperate regions that improve
pasture quality in both natural and cultivated mead-
ows. The species with the highest agricultural value
are T. alexandrinum,T. fragiferum,T. incarnatum T.
pratense and T. repens [15].
Choosing high-quality and suitable seed species
suitable for land and climate conditions is necessary
to obtain high yield efficiency. Human experts of-
ten control the grading of seeds according to quality
and type [16]. This process is susceptible to errors
and subjective because it depends on visual inspec-
tion. Choosing objective and quantitative methods to
classify seeds will satisfy both the producer and con-
sumer. For this purpose, many studies in the litera-
ture [17–19]. Recent studies have shown that state-of-
the-art solutions can successfully achieve the desired
output.
This study aims to classify five different Tri-
folium seeds depending on the features obtained by
using image processing and computer vision meth-
ods. There are many classification techniques in the
literature. These are divided into statistics-based,
rule-based, neural network-based, ensemble learning
and instance-based categories according to different
mathematical concepts used in classification. Naïve
Bayes, Decision Tree, Artificial Neural Networks
(ANN), Random Forest and k-Nearest-Neighbour(k-
NN) techniques, which are widely used in each cate-
gory, were used in this study. In addition, the data set
was trained with deep learning and results were ob-
tained. Thus, it was possible to compare two differ-
ent approaches. Section 2 gives detailed information
about the techniques.
2 Materials and Method
2.1 Image Acquisition
The computer vision process’s first stage is obtaining
the necessary digital images. The following steps’
success depends on the image set’s relevance and
quality. We have captured 1903 images of five differ-
ent species. A digital microscope acquired digital im-
ages of seed samples with a spatial resolution of 0.01
mm/pixel, a colour depth of 24 bits and a scanning
area not exceeding 1600 x 1200 pixels. The seeds
were placed on a whiteboard. Thus, the background
colour is different from the seed colours to ensure that
the segmentation process is more successful. A sam-
ple image of each Trifolium type is presented in Fig-
ure 1. The colour images were then converted to a
grey scale to be able to apply the next image process-
ing methods. A flowchart showing the order of each
operation is depicted in Figure 2.
(a) (b) (c)
(d) (e)
Figure 1: Sample images of different species of Tri-
folium (a) T. alexandrinum (b)T. fragiferum (c)T. in-
carnatum (d) T. pratense (e) T. repens
Figure 2: Flowchart of Trifolium seed classification
2.2 Image Segmentation
Image segmentation separates the object of interest
from other parts of the image [20]. There are var-
ious image segmentation methods, such as region-
based methods [21], morphological operation [20]
and thresholding [22]. The choice of the algorithm
used in segmentation usually depends on the prop-
erties of the problem solved. In this study, Otsu’s
threshold-based segmentation algorithm, which max-
imizes the inter-class variance parameter, was applied
to determine the threshold value.
Image segmentation identifies mutually exclusive
regions to find distinct elements that constitute an im-
age. These regions have either a common colour or
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
314
Volume 22, 2023
texture. The goal of segmentation is to find objects or
their boundaries. Segmentation plays a crucial role in
subsequent feature extraction and classification tasks.
There are two groups of segmentation methods—the
first one search for the edges, which display sharp dif-
ferences between the two contrasting regions. The
second one tries to discover related areas according
to defined similarity metrics. Thresholding is a sim-
ple but effective way of finding the contrast between
foreground and background objects. The basic idea
of thresholding is giving a specific intensity value for
the region with less than a predefined intensity value.
Pixels outside the area get a higher value. The thresh-
old value is decided depending on the outcome of his-
togram analysis. In some cases, the local threshold
method produces better results than the global thresh-
old method since the relationship between the pix-
els is unrelated to the location. Therefore, the lo-
cal method uses multiple threshold values in different
segments.
(a) (b)
Figure 3: Segmentation of Trifolium seeds (a) convex
hull (b) binary image after thresholding
2.3 Feature Extraction
A feature is defined as an essential part or aspect
of something. Feature extraction methods analyze
the entire or part of the image to determine the at-
tributes that can be stored as a quantitative value re-
quired by classification methods. Identifying more
relevant features to solve a specific problem can im-
prove the accuracy of classification models. Many
image-processing methods identify images or videos’
shape, colour, motion, and texture features. Table 1
presents the morphological features and their descrip-
tive statistics.
2.4 Classification Methods
Machine learning is based on the idea that the com-
puter can learn without being programmed explic-
itly. Machine learning requires the following tasks:
selection of sample data, pre-processing and trans-
forming data, model building and evaluation. Ma-
T.ale. T.fra. T. inc. T. pra. T. rep.
Count 346 410 348 407 392
Area
µ3.17 2.08 3.85 2.97 1.35
σ0.27 0.14 0.42 0.22 0.14
min 2.69 1.82 3.09 2.56 1.11
max 3.70 2.36 4.65 3.38 1.61
Perimeter
µ6.80 5.46 7.49 6.63 4.38
σ0.30 0.19 0.41 0.26 0.22
min 6.18 5.06 6.66 6.08 3.97
max 7.41 5.86 8.36 7.19 4.83
BBoxArea
µ4.16 2.72 5.07 3.97 1.76
σ0.38 0.21 0.55 0.32 0.19
min 3.42 2.29 3.90 3.32 1.40
max 5.05 3.25 6.24 4.72 2.18
ConvexArea
µ3.18 2.10 3.86 3.00 1.36
σ0.27 0.15 0.42 0.22 0.14
min 2.70 1.83 3.10 2.58 1.12
max 3.71 2.38 4.67 3.43 1.63
Diameter
µ2.01 1.63 2.21 1.94 1.31
σ0.09 0.06 0.12 0.07 0.07
min 1.85 1.52 1.98 1.81 1.19
max 2.17 1.73 2.43 2.07 1.43
F.Area
µ3.17 2.08 3.85 2.97 1.35
σ0.27 0.14 0.42 0.22 0.14
min 2.69 1.82 3.09 2.56 1.11
max 3.70 2.36 4.65 3.38 1.61
Major
µ2.37 1.80 2.62 2.29 1.40
σ0.14 0.07 0.16 0.13 0.08
min 2.06 1.59 2.25 1.93 1.22
max 2.73 2.03 3.03 2.63 1.68
Minor
µ1.70 1.48 1.87 1.67 1.24
σ0.10 0.06 0.13 0.09 0.07
min 1.44 1.28 1.54 1.38 1.04
max 1.96 1.62 2.21 1.95 1.40
Table 1: Descriptive statistics of Trifolium species
chine learning methods are generally categorized into
three main subgroups. These are supervised, unsu-
pervised and reinforcement learning methods. Super-
vised (predictive) learning methods use historical data
to predict future events. They are intended to estimate
class labels by using inputs (features). Classification,
Anomaly Detection and Regression belong to super-
vised learning. On the contrary, unsupervised (de-
scriptive) learning methods do not have class labels in
the data sets. These methods use similarity or distance
metrics to find groups of objects that exhibit common
behaviours. Clustering and Association Rule Min-
ing methods are two generally preferred unsupervised
methods. Lastly, the reinforcement learning method
determines its behaviour according to the reward and
punishment it receives in response to its actions. This
type of learning is widely preferred in digital games,
autonomous cars and robot programming.
The classification methods use some data (training
set) to determine the class boundaries. After the edges
are located, the class of an object whose label is un-
known is decided according to its position. This pro-
cess is defined as classification. The most common
classification algorithms are as follows: Decision
trees, Naïve Bayes, Logistic regression, k-Nearest
Neighbours(k-NN), Support Vector Machines(SVM)
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
315
Volume 22, 2023
and Artificial Neural Networks(ANN). In this study,
we will use four extensively preferred classification
methods.
2.4.1 Decision Trees and Random Forest
Decision trees are a non-parametric supervised learn-
ing method. Decision trees, which can process quan-
titative and categorical data, are widely used because
they are easy to build and interpret. In addition, they
are resistant to outliers. There are three types of nodes
in decision trees, root, leaf and internal. Decision tree
algorithms such as ID3, C4.5 and CART are based
on a divide-and-conquer approach. They start from
the root node and grow to the leaf node with a greedy
strategy. The leaf node at the end of the path start-
ing from the root determines the class label of an un-
labeled object. Decision trees are prone to overfit-
ting, mainly when the tree is fully-grown. A pruning
or random forest approach can be adopted to over-
come this problem. The pruning cut down less signif-
icant branches to create simpler trees. Random forest
builds multiple trees using sampling with replacement
methodology. If there is a conflict in the predictions
of the trees, the final class label is determined by eval-
uating the majority of the decisions made by each tree.
2.4.2 Naïve Bayes
The Naive Bayes method based on the Bayes the-
orem employs the assumption of independence be-
tween predictors. It calculates the probability of past
observations using training data set. Then it uses
probability values to estimate the likelihood of fu-
ture events. For each class, the Naive Bayes classifier
computes the posterior probability using Eq. 1. The
labels of the test data are determined as the label of
the class with the highest probability.
p(class |f eature) = p(feature |class)p(class)
p(feature)
(1)
• p(class |feature) : posterior probability of class
given evidence
• p(class) : prior probability of class
• p(feature |class) : prior probability of predictor
given class
• p(feature) : the prior probability of predictor
2.4.3 k-Nearest-Neighbour
Classification methods are divided into two cate-
gories as eager and lazy learners. Eager learners pro-
duce a prediction model using historical data before
classifying the new instance. Lazy learners do not
make a similar preparation before the prediction pro-
cess. Whenever a new unclassified object arrives, it
repeats the necessary steps defined by the algorithm
each time. k-Nearest-Neighbour (k-NN) is a lazy
learner. It finds the nearest kneighbours around the
unclassified object and decides the class label based
on the class labels of the neighbours. It determines
its closest neighbours using distance metrics. How-
ever, one of the most important disadvantages is the
determination of the kvalue that generates the best
prediction accuracy. For this purpose, a sequence of
experiments is carried out, and the kvalue that gives
the best result is chosen.
2.4.4 Artificial Neural Networks
Artificial neural networks, which are highly robust
and learn very quickly, simulate how the human brain
processes information. Artificial neural networks
are created by intensely interconnecting a series of
neuron-like simple units. Each unit converts the in-
put value it receives into an output. The generated
output is forwarded to the next layer. The output
value delivered from the last layer is used as a re-
sult. The connection between units is called edges,
which have weights initialized randomly. The learn-
ing process is defined as tuning these weights to en-
hance the accuracy of the network. There is a wide
variety of network types defined to handle different
problems. Feedforward, Radial Basis Function, Mul-
tilayer Perceptron, Convolutional and Recurrent Neu-
ral networks. Figure 4 illustrates the network archi-
tecture used in our study.
Figure 4: The Feedforward ANN structure
2.4.5 Convolutional Neural Networks
Deep learning is used to solve many different com-
plex problems in digital image processing. Deep
learning methods, such as Convolutional Neural Net-
works (CNNs), enhance the accuracy of classification
models by using a large data set and sufficient com-
puting power. CNNs consist of various layers: in-
put, convolution, pooling, fully connected, softmax
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
316
Volume 22, 2023
and output. In a deep learning architecture, the first
layer that receives raw pixels of input data learns how
to represent simple features. Each successive layer
learns the complex features of the previous layer as
they collect and recombine the features of the earlier
layers. Adding more layers to the network allows it
to handle higher dimensional data.
Although VGG was not the winner of the 2014
ImageNet Large-Scale Visual Recognition Challenge
(ILSVRC) competition, it has been used in many ap-
plications due to its modular and simpler architec-
ture [23]. Unlike other CNNs, VGG prefers smaller
kernel sizes, which helps increase its expressibility.
VGG architecture has several variations according to
the number of layers it contains. The smallest one
(VGG-11) has 11 layers and 133 million parameters.
The largest one (VGG-19) consists of 19 layers and
requires 144 million parameters to be tuned. The ar-
chitecture of the VGG is displayed in Figure 5.
Figure 5: The architecture of VGG [24]
2.5 Model Evaluation
At the final stage of the machine learning process,
each classification model should be evaluated with
test data they do not encounter during the building
phase. Thus, it can be foreseen how accurate they
can produce a prediction for objects whose label is
not specified. Besides, the best model can be distin-
guished among different ones. There are two widely
used evaluation methods.
2.5.1 Holdout Method
Before the model-building process, this method di-
vides the data into two disjoint sets. Generally, two-
thirds of the data is called training data used for model
building. The remaining portion is used for evalua-
tion purposes. If there is less training data, a model
with inadequate generalization capability can be de-
veloped. In this case, the accuracy of the test data may
not exceed a certain threshold. If the test data is in-
sufficient, an unreliable and unstable model emerges.
Repeating the holdout method with different training
and test data sets can reduce deviations caused by spe-
cific sample data. The overall error rate is computed
using the average error rates obtained for each data
set.
2.5.2 k-fold Cross-Validation
Unlike the Holdout method, the data set is divided into
kindependent sets of almost equal size. The testing
process is repeated ktimes in total. For each round,
k-1 sets are considered the training set and the model
is built using them. The model is tested with the ex-
cluded set that is different each time. The error rate is
calculated by dividing the total number of incorrectly
predicted objects by the total number of objects in the
data set. The advantage of this method is that it al-
lows testing of each object in the data set. Choosing k
is practically suitable between 5 and 10. In that case,
the test set consists of at least 10% of all data.
3 Results and Discussion
A digital image consists of a collection of pixels rep-
resenting an intensity value stored in a multidimen-
sional array. Naturally, raw image data is high di-
mensional. Therefore, feature selection and transfor-
mation play an essential role in the accuracy of image
classification applications. Irrelevant and redundant
attributes should be verified and removed from the
data set before the model building phase.
Raw features are converted to the following fea-
ture set:
• eccentricity: Eccentricity is the ratio of the focal
length of an ellipse to the major axis length. It
gets values in the range of zero and one. As the
value approaches zero, it is more like a circle.
• extent: It is computed as the ratio of the number
of pixels in the area to the number of pixels in the
total bounding box.
• shent: Entropy shows variability in data.
• solidity: It is expressed as the ratio of pixels in
the region to the pixels of the convex hull of the
image.
• minor_axis_length: The width of bounding box
• major_axis_length: The length of bounding box.
• maovermi: The value obtained by dividing the
major axis length by the minor axis length.
• are2per: It is computed by dividing the area into
the perimeter.
There are two different approaches to achieving fea-
ture selection. The first one calculates either correla-
tion or chi-squared values between features. Highly
correlated features are eliminated. The second ap-
proach uses a brute force search strategy to find the
optimal feature combination that results in the best ac-
curacy value. We utilized the first approach. Figure
6 presents the correlation matrix between features.
As seen from the correlation matrix, maovermi
attribute is highly correlated with eccentricity and
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
317
Volume 22, 2023
Figure 6: Correlation matrix of features
are2per (correlation value of maovermi between ec-
centricity and are2per is -0.98 and -0.82, respec-
tively). Therefore, maovermi is removed from the
data set. Then, four classification algorithms are ap-
plied. We use the scikit-learn Python library to run
each classifier for two different evaluation methods.
The accuracy percentage, weighted ROC area and F1
score values of each classification algorithm are pre-
sented in Table 2.
10-Fold Cross-Validation Holdout
Accuracy ROC F1 Accuracy ROC F1
Decision Tree 92.07%0.97 0.92 93.20%0.98 0.92
Random Forest 93.12%0.99 0.93 93.66%0.99 0.92
k-NN 92.28%0.99 0.92 92.73%0.99 0.93
Naive Bayes 91.96%0.99 0.90 92.12%0.99 0.92
ANN 93,17%∗0.99 0.93 94.59%∗0.95 0.99
Table 2: Performance of classifiers
Among the classification models produced, Naïve
Bayes has the lowest accuracy values for the 10-
fold cross-validation and holdout method. ANN has
the highest accuracy values for both methods. The
weighted ROC area and F1 score values of all models
are close to one, which indicates that each model con-
vincingly distinguishes objects of different classes.
According to two different evaluation methods, the
confusion matrices of ANN that have the highest ac-
curacy values are given in Table 3 and 4.
Actual Class
T.ale. T. fra. T. inc. T. pra. T. rep.
Predicted
T.alexandrinum 87 0 19 5 0
T. fragiferum 0 151 0 0 0
T. incarnatum 5 0 102 3 0
T. pratense 2 0 1 131 0
T. repens 0 0 0 0 141
Table 3: Confusion matrix of ANN for Holdout
Actual Class
T. ale. T. fra. T. inc. T. pra. T. rep.
Predicted
T. alexandrinum 272 0 52 22 0
T. fragiferum 0 410 0 0 0
T. incarnatum 29 0 316 3 0
T. pratense 22 0 2 383 0
T. repens 0 0 0 0 392
Table 4: Confusion matrix of ANN for 10-fold-Cross-
Validation
Actual Class
T.ale. T. fra. T. inc. T. pra. T. rep.
Predicted
T.alexandrinum 95 0 13 3 0
T. fragiferum 0 151 0 0 0
T. incarnatum 4 0 104 2 0
T. pratense 1 0 1 132 0
T. repens 0 0 0 0 141
Table 5: Confusion matrix of VGG19
All of the instances of T. fragiferum and T. repens
are correctly classified. However, prediction mod-
els incorrectly classify some instances that belong
to classes T. alexandrinum,T. incarnatum, and T.
pratense. We can conclude that T. fragiferum and
T. repens have more distinctive features than other
classes. The accuracy and loss of training and val-
idation sets for the VGG19 are shown in Figure 7.
The accuracy of both the training and validation data
sets converges to their final values at approximately
epoch 30.
Figure 7: Accuracy and loss of training and validation
sets
4 Conclusions
This study proposes a method that automates the clas-
sification process of Trifolium seed. An image data
set of five different species of Trifolium has been cre-
ated that can be used in future studies. After the nec-
essary prepossessing methods were applied to the im-
ages, they were forwarded to different classification
methods, including Decision tree, Naïve Bayes, k-
NN and ANN. The outcome of each classifier was
analyzed and discussed. It is realized that ANN pro-
duces the best performance values among the four
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
318
Volume 22, 2023
methods. Furthermore, the results demonstrate that
species of T. fragiferum and T. repens have distinct
quantitative features than the other ones. In addition
to the old-school methods mentioned above, we ob-
tained results by applying the VGG19 architecture to
the same data set. The performance of VGG19 is
slightly better than theirs. In conclusion, it has been
shown that traditional methods can produce predic-
tion models as accurate and reliable as deep learning
methods in cases where the dataset size is small and
computational power is low.
Acknowledgments: We would like to thank Re-
public of Turkey Ministry of Agriculture and Forestry
Variety Registration and Seed Certification Center for
providing the Trifolium seeds.
References:
[1] P. Ridao, M. Carreras, D. Ribas, and R. Gar-
cia, “Visual inspection of hydroelectric dams us-
ing an autonomous underwater vehicle,” Jour-
nal of Field Robotics, vol. 27, no. 6, pp. 759–
778, 2010.
[2] C. Kanellakis and G. Nikolakopoulos, “Survey
on computer vision for uavs: Current devel-
opments and trends,” Journal of Intelligent &
Robotic Systems, vol. 87, no. 1, pp. 141–168,
2017.
[3] J. Gao, Y. Yang, P. Lin, and D. S. Park, “Com-
puter vision in healthcare applications,” Journal
of Healthcare Engineering, vol. 2018, 2018.
[4] A. Esteva, A. Robicquet, B. Ramsundar,
V. Kuleshov, M. DePristo, K. Chou, C. Cui,
G. Corrado, S. Thrun, and J. Dean, “A guide to
deep learning in healthcare,” Nature Medicine,
vol. 25, no. 1, pp. 24–29, 2019.
[5] J. Ma, D.-W. Sun, J.-H. Qu, D. Liu, H. Pu,
W.-H. Gao, and X.-A. Zeng, “Applications of
computer vision for assessing quality of agri-
food products: A review of recent research ad-
vances,” Critical Reviews in Food Science and
Nutrition, vol. 56, no. 1, pp. 113–127, 2016.
[6] A. Taheri-Garavand, S. Fatahi, M. Omid, and
Y. Makino, “Meat quality evaluation based on
computer vision technique: A review,” Meat
science, 2019.
[7] R. Brunelli and T. Poggio, “Face recogni-
tion through geometrical features,” in European
Conference on Computer Vision, pp. 792–800,
1992.
[8] T. U. Rehman, M. S. Mahmud, Y. K. Chang,
J. Jin, and J. Shin, “Current and future applica-
tions of statistical machine learning algorithms
for agricultural machine vision systems,” Com-
puters and Electronics in Agriculture, vol. 156,
pp. 585–605, 2019.
[9] W. Qian, Y. Huang, Q. Liu, W. Fan, Z. Sun,
H. Dong, F. Wan, and X. Qiao, “Uav and a deep
convolutional neural network for monitoring in-
vasive alien plants in the wild,” Computers and
Electronics in Agriculture, vol. 174, p. 105519,
2020.
[10] E. Kamir, F. Waldner, and Z. Hochman, “Es-
timating wheat yields in australia using cli-
mate records, satellite image time series and
machine learning methods,” ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 160,
pp. 124–135, 2020.
[11] M. Zohary, D. Heller, et al.,The genus Tri-
folium. Israel Academy of Sciences and Human-
ities, 1984.
[12] L. Zoric, L. Merkulov, J. Lukovic, and P. Boza,
“Comparative analysis of qualitative anatomi-
cal characters of trifolium l.(fabaceae) and their
taxonomic implications: preliminary results,”
Plant Systematics and Evolution, vol. 298, no. 1,
pp. 205–219, 2012.
[13] W. Oleszek and A. Stochmal, “Triterpene
saponins and flavonoids in the seeds of tri-
folium species,” Phytochemistry, vol. 61, no. 2,
pp. 165–170, 2002.
[14] J. Russell and H. Webb, “Climatic range of
grasses and legumes used in pastures: Results
of a survey,” The Journal of the Australian Insti-
tute of Agricultural Science, vol. 3, p. 156–166,
1976.
[15] J. Kolodziejczyk-Czepas, “Trifolium species–
the latest findings on chemical profile, eth-
nomedicinal use and pharmacological proper-
ties,” Journal of Pharmacy and Pharmacology,
vol. 68, no. 7, pp. 845–861, 2016.
[16] P. Dubosclard, S. Larnier, H. Konik, A. Herbu-
lot, and M. Devy, “Automatic method for vi-
sual grading of seed food products,” in Interna-
tional Conference Image Analysis and Recogni-
tion, pp. 485–495, 2014.
[17] F. Guevara-Hernandez and J. G. Gil, “A ma-
chine vision system for classification of wheat
and barley grain kernels,” Spanish Journal of
Agricultural Research, vol. 9, no. 3, pp. 672–
680, 2011.
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
319
Volume 22, 2023
[18] E. O. Güneş, S. Aygün, M. Kırcı, A. Kalateh,
and Y. Çakır, “Determination of the varieties
and characteristics of wheat seeds grown in
turkey using image processing techniques,” in
2014 The Third International Conference on
Agro-Geoinformatics, pp. 1–4, 2014.
[19] A. Bagheri, L. Eghbali, and R. S. Haghighi,
“Seed classification of three species of amaranth
(amaranthus spp.) using artificial neural network
and canonical discriminant analysis,” The Jour-
nal of Agricultural Science, vol. 157, no. 4,
pp. 333–341, 2019.
[20] R. C. Gonzalez and R. E. Woods, Digital Im-
age Processing. USA: Addison-Wesley Long-
man Publishing Co., Inc., 2nd ed., 2001.
[21] C. Huang, Q. Liu, and X. Li, “Color image seg-
mentation by seeded region growing and region
merging,” in 2010 Seventh International Con-
ference on Fuzzy Systems and Knowledge Dis-
covery, vol. 2, pp. 533–536, 2010.
[22] N. Otsu, “A threshold selection method from
gray-level histograms,” IEEE Transactions on
Systems, Man, and Cybernetics, vol. 9, no. 1,
pp. 62–66, 1979.
[23] K. Simonyan and A. Zisserman, “Very deep con-
volutional networks for large-scale image recog-
nition,” arXiv preprint arXiv:1409.1556, 2014.
[24] Y. Altuntaş, Z. Cömert, and A. F. Kocamaz,
“Identification of haploid and diploid maize
seeds using convolutional neural networks and
a transfer learning approach,” Computers and
Electronics in Agriculture, vol. 163, p. 104874,
2019.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflict of interest to
declare that is relevant to the content of this
article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2023.22.34
Recep Eryigit, Yilmaz Ar, Bulent Tugrul
E-ISSN: 2224-2678
320
Volume 22, 2023
