A Computer Vision-Based Framework for Multilevel Egg Fertility Classification Using L*a*b* Color Segmentation, Wavelet Texture Features, and Support Vector Machine

Vol. 8, No. 5, October 2026

6857
Journal of Information Systems and Informatics | ISSN: 2656-5935 | e-ISSN: 2656-4882 | pp. 6857-6889
Published by Asosiasi Doktor Sistem Informasi Indonesia
A Computer Vision-Based Framework for Multilevel Egg Fertility
Classification Using Lab* Color Segmentation, Wavelet Texture
Features, and Support Vector Machine

Sriyadi1*, Jufriadif Na'am2, Mochamad Wahyudi3, Sumanto4

1,2Doctoral Program of Informatics, Universitas Nusa Mandiri, Jakarta, Indonesia
1Department of Information System, Universitas Bina Sarana Informatika, Jakarta, Indonesia
3,4Master of Information Technology Program, Universitas Bina Sarana Informatika, Jakarta, Indonesia

Received:
June 30, 2026
Revised:
September 5, 2026
Accepted:
October 1, 2026
Published:
October 24, 2026

Corresponding Author:
Author Name*:
Sriyadi
Email*:
[email protected]

DOI:
10.63158/journalisi.v8i5.1882

© 2026 Journal of
Information Systems and
Informatics. This open
access article is distributed
under a (CC-BY License)

Abstract. Egg fertility assessment is essential for improving
hatchery efficiency, reducing incubation costs, and minimizing
resource waste caused by infertile eggs. This study aims to develop
a computer vision-based framework for three-level egg fertility
classification, namely Highly Fertile, Fertile, and Infertile. A balanced
dataset consisting of 1,500 manually collected candling images (500
images per class) was validated by three poultry experts. The
proposed method involves Lab* color-space transformation, K-
Means clustering for region-of-interest segmentation, Discrete
Wavelet Transform (DWT) using the LL sub-band for texture feature
extraction, and multiclass Support Vector Machine (SVM)
classification. The experimental results show that the proposed
model achieves 90.9% training accuracy, with precision, recall, and
F1-score of 92.0%, 90.9%, and 90.5%, respectively. For the
independent testing dataset, the model achieves 89.7% accuracy
and 89.4% F1-score, indicating good generalization capability with
only a 1.2% difference between training and testing accuracy. The
novelty of this research is the development of a three-level egg
fertility classification framework that extends conventional binary
fertility detection by integrating Lab*-based segmentation, LL-
wavelet statistical texture features, and multiclass SVM
classification.

Keywords: egg fertility classification; egg candling; computer
vision; image classification; machine learning; agricultural
automation; image processing.

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6858

1. INTRODUCTION

The fertility of eggs and their classification are crucial aspects of the poultry industry
because they directly affect hatchability, production efficiency, and overall egg product
quality. The early detection of egg fertility enables poultry farmers to separate fertile
from infertile eggs before incubation, thereby reducing resource waste and improving
productivity [1], [2], [3], [4], [5]. Non-destructive testing techniques have received
considerable attention in monitoring the quality of food and agricultural products, with
hyperspectral imaging being an important technique because it provides both spatial and
spectral information simultaneously [6]. Advances in machine learning techniques for
rapid analysis and higher classification accuracy have further increased the potential of
these methods for food-related applications, including egg classification [6]. In this
context, the development of an automated system based on image processing and
machine learning is essential to replace manual inspection methods, which are often time-
consuming, subjective, and prone to human error [7]. For instance, eggshell crack detection
has been considered important for improving fresh egg consumption, breeding egg
hatchability, and overall egg product quality [8]. Therefore, research on egg fertility
classification using intelligent computational approaches is highly relevant for supporting
modern hatchery industries that require efficient, accurate, and objective fertility
assessment systems.

Although various methods have been developed for egg quality detection and
classification, several significant challenges remain and require further investigation. One
of the main challenges is the availability of adequate and high-quality datasets. In many
related application domains, the absence of standardized public datasets becomes a
serious limitation; for example, in automotive surface defect detection, no public dataset
is available for classification, resulting in limited data samples for model development [9].
This dataset issue is also relevant to egg fertility classification, where variations in lighting
conditions, egg position, and embryonic development stages introduce complexities that
are not fully represented in existing datasets. A previous study reported that fertile and
infertile egg classification achieved only 84% accuracy on the fourth day of incubation
[6], indicating that existing methods do not consistently provide satisfactory results. In
related domains, several classification approaches have also shown similar limitations; for
example, the region growing algorithm achieved only 78% accuracy in follicle recognition

Vol. 8, No. 5, October 2026

6859 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
[10], SVM-based mammogram classification achieved 80%–83% accuracy [11], and ANN and
PCA-ANN approaches for brain tumor image classification produced 75%–90% accuracy
[12]. Furthermore, classification using dragonfly clusters with KNN obtained the lowest
accuracy of 75.5%, whereas firefly clusters with NBC achieved 75.75% [13]. Similarly,
Wavelet scattering with SVM as weak classifiers achieved average accuracies of only 61%
and 66%, respectively [14]. These findings indicate that single-method approaches are
often insufficient for handling the complexity of biological image data, and a more
comprehensive integration of methods is required. Based on previous studies, the
classification performance on egg objects and egg quality characteristics still varies
considerably, particularly in studies reporting accuracies below 85%. This indicates that
previous methods have limitations in feature representation, classification techniques, and
model capability in handling variations in egg characteristics. Relevant previous studies
are summarized in Table 1.

Table 1. Summary of Previous Studies
No
Researcher
(Year)
Number of
Classes
Segmentation
Methods
Feature
Extraction
Methods
Classification
Methods
Accuracy
1
Saha &
Manickavasagan
(2021) [15]
2: fertile,
infertile
K-means
clustering
— SVM
84% on
day 4
2
Shi et al. (2022)
[9]
2: cracked,
intact
—
Wavelet
features
SVM, LDA,
Decision Tree
82.5%
3
Elibol et al.
(2020)[16]
4: extra
large, large,
medium,
small
— DWT SVM, LBNC 80.4%
4
Adegbenjo et al.
(2020)[17]
2–3: fertile,
infertile,
dead
embryo
Conventional
candling /
machine
vision
Spectral
and
textural
features
Various
conventional
methods
<85%

Therefore, in this study, K-means clustering was employed as an initial segmentation
method to separate the region of interest from egg images prior to feature extraction.
This segmentation stage is important to ensure that the features extracted in the

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6860
subsequent stage focus more specifically on image regions that are relevant to egg
fertility indicators.

K-means clustering has several fundamental advantages that make it suitable for
segmentation in egg fertility classification. Computationally, K-means is a simple, efficient,
and easy-to-implement algorithm for partitioning data into groups based on pixel
similarity [11],[19] In egg image applications, segmentation is a critical stage because egg
images captured using candling techniques may contain complex intensity variations
among the embryo region, air cell, yolk, and eggshell. K-means can automatically separate
the region of interest, particularly the embryo-related area as the main fertility indicator,
from the background and other egg components without prior knowledge or initial labels
[8], [18]. Another advantage of K-means is its ability to handle significant pixel differences
between the target region and the surrounding areas, allowing the clustering process to
segment the desired region more accurately [18], [19], [20], [21], [22]. In the context of egg
fertility classification, the intensity differences between fertile eggs with developing
embryos and infertile eggs without visible embryo development create distinct pixel
patterns. These differences can be utilized by K -means to produce more precise
segmentation results. In addition, histogram-based centroid initialization can be used to
reduce the randomness of initial centroid selection, which often causes convergence to
local solutions, thereby improving the consistency of segmentation results [18]. K-means
has also been proven effective in medical and biological image segmentation tasks that
share similar characteristics with egg images, such as tumor segmentation in MRI images
[18], [19] and follicle detection in ovarian ultrasound images [8]. Therefore, the ability of K-
means to separate relevant image regions makes it a strong foundation for the
subsequent feature extraction stage [19].

However, K-means segmentation alone is insufficient to produce accurate classification
results because the segmented output is still an image that requires a discriminative
numerical feature representation to distinguish fertility levels. During this stage, the
discrete Wavelet transform (DWT) plays a crucial role as a bridge between segmentation
and classification [12], [23], [24], [25], [26]. The DWT can extract multi-resolution information
from images by decomposing them into approximation and detail components, including
horizontal, vertical, and diagonal details at different levels. This enables the extraction of
textural, edge, and subtle pattern characteristics that may not be visible at a single

Vol. 8, No. 5, October 2026

6861 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
resolution [12], [23]. In the context of egg fertility classification, differences between fertile
and infertile eggs often appear in fine textural patterns and micro-intensity variations,
which can be captured more effectively through a multi-scale analysis provided by the
Wavelet transform [9], [12]. The DWT also has a useful time–frequency localization
property, meaning that the Wavelet can simultaneously identify where and at what
frequency significant changes occur in an image [19]. The integration of K-means and
wavelets is therefore important because K-means first isolates the relevant region of
interest, and the DWT then extracts discriminative features from that region with less
contamination from irrelevant background pixels. This combination produces cleaner and
more informative features than applying the Wavelet directly to the raw image [11], [18],
[19], [22], [27], [28]. Moreover, removing the Wavelet transform module from a classification
pipeline has been shown to significantly reduce the accuracy, confirming the essential
role of the Wavelet in feature extraction [9]. The resulting Wavelet features then serve as
suitable inputs for a support vector machine (SVM), which is known for its high
generalization capability in high-dimensional feature spaces and limited datasets [9], [15].

SVM is an efficient machine-learning algorithm for decision-making and pattern
recognition, with a key advantage in its regularization parameter, which helps reduce the
risk of overfitting [18]. SVM performs well in small-sample classification and high-
dimensional feature spaces, where its computational complexity is not directly dependent
on the dimensionality of the input space. In addition, multi-class SVM is considered robust
in handling overfitting problems [29], [30], [31]. In the context of limited datasets, SVM is
suitable because its computation depends mainly on the number of support vectors rather
than the total number of samples [9]. SVM has also been shown to produce more accurate
classification results than other methods, such as neural network learning and k-nearest
neighbors, in several biomedical image classification applications [10]. The combination of
these three methods—K-means for segmentation, Wavelet for feature extraction, and SVM
for classification—has demonstrated strong effectiveness in various domains, including
skin disease identification with an accuracy of 97.1% [18], eggshell crack detection with an
accuracy above 99% [8], and medical image classification [23], [32].

Previous studies have demonstrated the potential of computer vision and machine
learning techniques for automated egg fertility assessment. However, most existing
studies have primarily focused on binary classification between fertile and infertile eggs

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6862
rather than assessing fertility levels. Pham (2023) [6] investigated fertile and infertile egg
classification during incubation and achieved 84% accuracy; however, the study was
limited to two-class classification. Other studies have applied image processing, machine
learning, and deep learning methods for egg fertility detection and egg quality assessment
[7], [8], [9], [10] but these approaches mainly addressed binary decisions or general quality
evaluation rather than fertility level grading. For example, Saha and Manickavasagan [15]
investigated fertile and infertile egg classification using SVM and achieved 84% accuracy
on day 4 of incubation. Similarly, other machine vision-based approaches have mainly
focused on binary classification or general egg quality assessment with limited
classification levels. These findings indicate that existing approaches still have limitations
in representing different fertility conditions and developmental stages required for
multilevel egg fertility assessment.

Although previous studies have demonstrated the potential of computer vision and
machine learning for egg fertility assessment, most existing approaches are still limited
to binary classification between fertile and infertile eggs. However, binary classification
provides only a general fertility decision and cannot represent different fertility
conditions and developmental stages that are important for more precise hatchery
management and decision-making. Therefore, this research proposes a three-level egg
fertility classification framework consisting of highly fertile, fertile, and infertile
categories.

The novelty of this research lies in developing a three-level egg fertility classification
framework that integrates La-based K-Means segmentation, Wavelet-based texture
feature extraction, and multiclass Support Vector Machine classification. In this
framework, La-based K-Means is used to extract relevant egg regions, Wavelet
transformation is applied to obtain discriminative texture features, and multiclass SVM is
employed to classify eggs into highly fertile, fertile, and infertile categories. Unlike
previous binary fertility detection approaches, the proposed framework provides a more
detailed fertility assessment for objective hatchery screening. The objectives of this
research are to develop an image segmentation approach for extracting relevant egg
regions, extract discriminative texture features using Wavelet transformation, and
evaluate the performance of multiclass SVM for three-level egg fertility classification.

Vol. 8, No. 5, October 2026

6863 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
2. METHODS

The overall workflow consists of five main stages: dataset collection and validation, data
splitting, Lab*-based K-Means segmentation, Wavelet-based feature extraction, and
multiclass SVM classification. The complete workflow of the proposed framework is
illustrated in Figure 1, the research process begins with egg image collection and expert-
based label validation to ensure the quality and correctness of the data used. The dataset
is then divided into training and testing datasets so that model development and
evaluation can be carried out proportionally. In the next stage, egg images are processed
using K-Means to obtain a relevant region of interest, and image features are extracted
using a Wavelet transform to represent texture and intensity patterns more
discriminatively. The extracted features are subsequently used as inputs for a support
vector machine to classify the eggs into three categories: highly fertile, fertile, and
infertile. Through this workflow, the proposed method is expected to provide a more
systematic and consistent classification process to support image-based egg fertility
analysis.

Figure 1. General Research Flowchart of the Proposed Multilevel Egg Fertility
Classification Framework

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6864

Figure 2. Research Stages

This study proposes an image-based egg fertility level classification framework by
combining K-means segmentation, Wavelet feature extraction, and support vector
machine (SVM) classification. The research method consists of five main stages: dataset
collection and validation, data splitting, image segmentation using K-means, feature
extraction using wavelets, and classification using SVM. This workflow is designed to
ensure that the classification process is carried out systematically, starting from the use
of an expert-validated dataset and culminating in the prediction of egg fertility classes.

Vol. 8, No. 5, October 2026

6865 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
2.1. Dataset Collection and Validation
The first stage involves the collection and validation of the egg image dataset. The dataset
was collected independently, considering image acquisition conditions such as lighting,
egg position, camera distance, and image background. These acquisition conditions were
controlled to ensure that the captured images had relatively consistent quality and could
be used effectively in the image processing stage.

The dataset used in this study comprised 1,500 egg images divided into three categories:
highly fertile, fertile, and infertile eggs. Each category comprised 500 images, resulting in
balanced class distribution. After image acquisition, each image was labeled according to
its fertility. The image labels were validated by three poultry experts based on the visual
fertility indicators observed during candling. In cases of disagreement, the final label was
determined by expert consensus. Expert validation is important because the label quality
directly affects the success of model training and testing.

Table 2. Dataset Distribution
No Egg Fertility Category
Number of
Images
Description
1 Highly Fertile 500 Validated by experts
2 Fertile 500 Validated by experts
3 Infertile 500 Validated by experts
Total 1500 Balanced dataset

The total number of images can be expressed using Equation (1).

𝑁=∑ 𝑁!"!#$ (1)

where 𝑁represents the total number of images, 𝐶represents the number of classes, and
𝑁! represents the number of images in the 𝑖-th class. In this study, three classes were
used, namely Highly Fertile, Fertile, and Infertile. Therefore, the total number of images
is calculated as follows:
N = 500 + 500 + 500 = 1500
C = 3

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6866
2.2. Data Splitting
The second stage involved data splitting. The validated dataset was divided into training
and testing datasets using an 80:20 ratio, with 80% of the images used for training and
20% for testing. The training data were used to build the classification model, whereas
the testing data were used to evaluate the final performance of the model on new data
that had not been used during the training process. Data splitting was performed
proportionally for each class to prevent the model from being biased toward a particular
fertility category. Therefore, each class had the same data distribution for training and
testing. Of the total of 1,500 images, 1,200 images were used as training data and 300
images were used as testing data. Each class consisted of 500 images, with 400 images
used for training and 100 images used for testing. The data division equation can be
expressed as follows:
𝑁%&’!(=0.80×𝑁
𝑁%)*%=0.20×𝑁

With a total dataset of 1,500 images, the following results were obtained:

𝑁%&’!(=0.80×1500=1200
𝑁%)*%=0.20×1500=300

For each class containing 500 images, the data distribution per class is:

𝑁%&’!(+,’=0.80×500=400
𝑁%)*%+,’=0.20×500=100

2.3. Image Segmentation Using K-means
The third step involved image segmentation using the K-means clustering algorithm.
Segmentation was performed to isolate the important parts of the egg image or regions
of interest (ROI)from other irrelevant areas, such as the background, shadows, and noise
[33]. In egg fertility classification, not all parts of the image contribute equally. Therefore,
segmentation is necessary so that the feature extraction process can focus more on
image areas related to egg fertility indicators. In this study, the K-means algorithm was
applied using the Lab* color space to improve segmentation quality. The Lab* color
space consists of three components: L* represents lightness, while a* and b* represent
chromatic information. This representation separates brightness information from color

Vol. 8, No. 5, October 2026

6867 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
components, allowing pixel differences to be represented more effectively during the
clustering process. The Lab* color space was chosen because it separates the lightness
(L*) component from the a* and b* color components, allowing the differences in color
characteristics within the egg images to be represented more clearly. The use of the
Lab* color space helps K-means cluster pixels based on color similarity more stably,
especially in images with varying lighting conditions. Thus, Lab* -based K-means is
expected to produce cleaner and more relevant ROI segmentation.

In this study, RGB images were first converted to the L* a* b* color space. Each pixel is
represented as a vector as shown in the equation (2) Each pixel is represented as a three-
dimensional feature vector consisting of L*, a*, and b* values, which are used as input
features for the K-Means clustering algorithm:

𝑥!=[𝐿!∗,𝑎!∗,𝑏!∗ (2)

K-means is used to group pixels into several clusters by minimizing the distance between
the pixels and centroids. The K-means objective function is expressed in Equation (3) as
follows:
𝐽=7 ∑ ∥.!∈""
0
1#$
𝑥!−𝜇1∥2 (3)

where 𝐽represents the objective function, 𝐾represents the number of clusters,
𝐶1represents the 𝑗-th cluster, 𝑥!represents the 𝑖-th pixel, and 𝜇1represents the centroid
of the 𝑗-th cluster. The distance between each pixel and the centroid is calculated using
Euclidean distance, as shown in Equation (4).

𝑑(𝑥!,𝜇1)=@A (34#$𝑥!4−𝜇14)2 (4)

where 𝑀represents the number of pixel features, 𝑥!4represents the value of the 𝑚-th
feature of the 𝑖-th pixel, and 𝜇14represents the value of the 𝑚-th feature of the 𝑗-th
centroid. After each pixel is assigned to the nearest cluster, the centroid is updated by
calculating the average of all pixels within the corresponding cluster, as shown in
Equation (5):
𝜇1=$∣""∣∑ 𝑥!.!∈"" (5)

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6868
where ∣𝐶1∣represents the number of pixels in the 𝑗-th cluster. The final output of this
stage is a segmented image that represents the region of interest (ROI), which is then
used as input for the Wavelet feature extraction process.

The clustering process is performed iteratively by assigning each pixel to the nearest
centroid, recalculating centroid values, and repeating the process until the centroid values
converge. The final cluster representing the relevant egg region is selected as the region
of interest (ROI) for further feature extraction.

2.4. Feature Extraction Using Wavelets
The fourth step is feature extraction using the Wavelet transform, specifically the
discrete Wavelet transform (DWT). After the egg image has been segmented, the ROI
image must be converted into a numerical representation for use by the classification
algorithm [34]. Wavelets are used because they can analyze images in a multiresolution
domain, thereby capturing information about the texture, edges, and intensity patterns
at various levels of detail. In one-level two-dimensional DWT decomposition, the ROI
image 𝐼(𝑥,𝑦)is decomposed into four sub-bands, as shown in Equation (6):

𝐼(𝑥,𝑦)→{𝐿𝐿,𝐿𝐻,𝐻𝐿,𝐻𝐻} (6)

The 𝐿𝐿sub-band represents the approximation component or low-frequency information,
while 𝐿𝐻, 𝐻𝐿, and 𝐻𝐻represent the horizontal, vertical, and diagonal detail components,
respectively. In this study, the LL sub-band was selected because it retains the main
structural and intensity information of the egg image. This component provides a more
stable representation of fertility-related patterns while reducing the influence of high-
frequency noise. In general, the Wavelet Transform can be expressed by the following
equation:
𝑊6(𝑎,𝑏)=$√∣’∣∫𝑓(𝑡)𝜓898 P%9:’Q𝑑𝑡 (7)

where 𝑊6(𝑎,𝑏) represents the Wavelet coefficient, 𝑓(𝑡)represents the input signal,
𝜓(𝑡)represents the mother wavelet, 𝑎represents the scale parameter, and 𝑏represents
the translation parameter.

Vol. 8, No. 5, October 2026

6869 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
Statistical features can be calculated from the Wavelet coefficients. Several features
used in this study include the mean, standard deviation, energy, and entropy. The mean
value is calculated using Equation (8).

𝑀𝑒𝑎𝑛=$(∑ 𝑐!(!#$ (8)

The standard deviation was calculated using Equation (9).

𝑆𝑡𝑑=@$(A ((!#$𝑐!−𝑐¯)2 (9)

Energy is calculated using Equation (10).

𝐸𝑛𝑒𝑟𝑔𝑦=A 𝑐!2(!#$ (10)

Entropy is calculated using Equation (11).

𝐸𝑛𝑡𝑟𝑜𝑝𝑦=−∑ 𝑝!(!#$ log2(𝑝!) (11)

where 𝑐!represents the 𝑖-th Wavelet coefficient, 𝑛represents the number of coefficients,
𝑐¯represents the mean of the Wavelet coefficients, and 𝑝!represents the probability of
occurrence of the 𝑖-th coefficient value. The features extracted from each sub-band are
then combined into a feature vector, as expressed in Equation (12).

𝐹=[𝑓$,𝑓2,𝑓;,...,𝑓(] (12)

This feature vector serves as a numerical representation of the egg image and is used
as input in the classification stage using an SVM.

2.5. Classification Using Support Vector Machines
The fifth step involves classifying egg fertility levels using a Support Vector Machine
(SVM). The feature vectors extracted via the Wavelet method are used as inputs for the
SVM model to classify egg images into three classes: highly fertile, fertile, and infertile.
The SVM works by finding an optimal hyperplane that can separate data across classes

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6870
with the maximum margin. For binary classification, the SVM decision function can be
expressed as shown in Equation (13)[18]:

𝑓(𝑥)=𝑤<𝑥+𝑏 (13)

Class predictions were determined based on

𝑦=𝑠𝑖𝑔𝑛(𝑤<𝑥+𝑏)

where 𝑥represents the input feature vector, 𝑤represents the hyperplane weight vector,
𝑏represents the bias, and 𝑦represents the class label.The optimization objective of SVM
is to minimize the following function:
min12∥𝑤∥2

subject to:
𝑦!(𝑤<𝑥!+𝑏)≥1

For data that cannot be linearly separated, soft-margin SVM is used:

min12∥𝑤∥2+𝐶7𝜉!
(
!#$

subject to:
𝑦!(𝑤<𝑥!+𝑏)≥1−𝜉!
𝜉!≥0

where 𝐶represents the regularization parameter and 𝜉!represents the slack variable.
Because this study uses three classes, an SVM is applied using a multiclass approach. One
approach that can be used is one-vs-rest, in which each class is compared with the
combination of the remaining classes. In the One-vs-Rest strategy, each SVM classifier is
trained to distinguish one fertility category from the other two categories. During
prediction, the decision scores from all classifiers are compared, and the class with the
highest decision value is selected as the final prediction. The number of SVM models in
the one-vs-rest approach is equal to the number of classes:
𝑀=𝐶

Vol. 8, No. 5, October 2026

6871 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
Since there are three classes in this study, then:

𝑀=3
The model that was created is:

𝑆𝑉𝑀$: Highly Fertile 𝑣𝑠. 𝑂𝑡ℎ𝑒𝑟𝑠
𝑆𝑉𝑀2:𝐹𝑒𝑟𝑡𝑖𝑙𝑒 𝑣𝑠.Other𝑠
𝑆𝑉𝑀;:𝐼𝑛𝑓𝑒𝑟𝑡𝑖𝑙𝑒 𝑣𝑠. 𝑂𝑡ℎ𝑒𝑟𝑠

The final class is determined based on the maximum value of the decision function:

𝑦w=argmax=#$
"
𝑓=(𝑥)

where 𝑦wrepresents the predicted class, 𝐶represents the number of classes, and
𝑓=(𝑥)represents the decision function value for the 𝑘-th class. Model evaluation was
performed using testing data that were not used during the training process. The
evaluation metrics used in this study include accuracy, precision, recall, and the F1-score.
Accuracy is calculated using Equation (14).

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦= <>?<@<>?<@?A>?A@ (14)

Precision is calculated using Equation (15).

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛=<><>?A> (15)

Recall is calculated using Equation (16).

𝑅𝑒𝑐𝑎𝑙𝑙=<><>?A@ (16)

The F1 score was calculated using the following equation (17):

𝐹1-𝑠𝑐𝑜𝑟𝑒=2×>&)+!!B(×D)+’,,>&)+!!B(?D)+’,, (17)

where 𝑇𝑃represents true positive, 𝑇𝑁represents true negative, 𝐹𝑃represents false
positive, and 𝐹𝑁represents false negative.

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6872
3. RESULTS AND DISCUSSION

This section presents the results of each research stage in sequence, starting with the
distribution of the dataset, segmentation results using K-means, feature extraction using
wavelets, classification using support vector machines, and ending with the evaluation of
model performance. The results are presented in accordance with the research
methodology to ensure that the relationships between each stage are clear and
systematic.

3.1. Dataset Distribution
The dataset used in this study consists of 1,500 egg images that have been validated by
experts. The dataset is divided into three fertility categories: highly fertile, fertile, and
infertile. Each category contains the same number of images (500). After the validation
process, the dataset was split into training and test sets at an 80% to 20% ratio.

Table 3. Distribution of the Research Dataset
Category Total Images Training Testing
Highly Fertile 500 400 100
Fertile 500 400 100
Infertile 500 400 100
Total 1500 1200 300

The data distribution in Table 3 shows that the number of images in each class is balanced.
This data balance is important for reducing potential model bias toward any one fertility
class. With this distribution, the model has sufficient data for training, can still be validated
during development, and can be tested using data that were not used during training.

3.2. K-means Clustering
Image segmentation in this study was performed using the K-means algorithm to separate
the region of interest (ROI) from egg images, allowing the analysis process to focus on
areas directly related to fertility indicators. The segmentation stage is important because
the acquired egg images may contain not only the main object but also background
elements, shadows, light reflections, and intensity variations that can interfere with the

Vol. 8, No. 5, October 2026

6873 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
feature extraction process. To obtain more optimal segmentation results, K-means was
applied in the Lab* color space. This color space was selected because it separates the
lightness component (L*) from the color components (a* and b*), allowing visual differences
in egg images to be represented more clearly than when using the RGB color space. The
separation between lightness and color information helps the K-means algorithm group
pixels more stably based on color similarity, particularly in images with lighting variations.

Using this approach, pixels with similar characteristics can be grouped into specific
clusters, whereas irrelevant regions can be separated from the main object. As shown in
Figure 2, the segmentation results obtained using Lab*-based K-means indicate that the
main egg region can be separated from the background and other unnecessary image
components. This result serves as an important initial stage before feature extraction
because the quality of the Wavelet features is strongly dependent on the image region
used as input. If the processed image still contains considerable noise, shadows, or
background information, the extracted features may become less representative of egg
fertility characteristics. Therefore, the use of Lab*-based K-means segmentation helps
produce cleaner, more focused, and more relevant input images for the subsequent
Wavelet feature extraction stage.

Figure 3. Segmentation Results Using K-means

As shown in Figure 3, the segmentation results indicate that lab-based K-means is
capable of generating more precise ROIs for each egg fertility category. These results

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6874
suggest that the segmentation stage has provided more relevant image input for the
subsequent Wavelet feature extraction process.

Figure 4. Segmentation Results Using K-means-Lab

3.3. Wavelet Feature Extraction
After K-means segmentation, the egg images underwent Wavelet transformation to obtain
a more informative numerical feature representation. As shown in Figure 4, the Wavelet
processing results for the three egg categories reveal distinct visual characterist ics
between the infertile, fertile, and highly fertile classes. In the unfertilized class, the image
appears brighter and relatively homogeneous, whereas in the fertilized class, dark areas
with certain variations in intensity begin to appear. The highly fertilized class shows more
complex textural patterns and intensity distributions. These differences indicate that the
Wavelet transform can capture important information in the form of intensity patterns,
textures, and changes in the visual structure of egg images. The results of the Wavelet
processing are shown in Figure 4.

Figure 5. Feature Extraction Results Using Wavelet Transform (LL Sub-band)

Vol. 8, No. 5, October 2026

6875 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
The results of the Wavelet feature extraction serve as a crucial foundation for the
classification process because the image is no longer represented merely as a visual image
but is transformed into numerical feature values, such as the mean, standard deviation,
energy, and entropy. These features are used to describe the characteristics of each egg
fertility class in a more quantifiable manner. With this feature representation, the SVM
classification model can learn the differences in patterns between clas ses more
systematically. Therefore, the Wavelet feature extraction stage acts as a bridge between
the image segmentation results and the classification process by providing the
discriminative information needed to distinguish between infertile, fertile, and highly
fertile categories. The results of each feature extraction, averaged across all images, are
presented in Table 4.
Table 4. Wavelet Feature Extraction Results
Category
Mean
(Average)
Standard Deviation
(Avg)
Energy (Average)
10¹⁰
Entropy
(Average)
Fertile 116.1557 22.7301 3.5097 5.1994
Infertile 116.7357 25.2094 4.1353 5.1276
Highly
Fertile
118.7481 21.2167 4.6269 4.8324

Based on the Wavelet feature extraction results on the LL sub-band, each egg fertility
class exhibited different feature values. The Fertile class attained a mean value of 116.1557,
a standard deviation of 22.7301, an energy value of 3.5097 × 10¹⁰, and an entropy value of
5.1994. The higher entropy value in the Fertile class indicated that images in this category
contained a relatively greater variation of information or textural irregularity compared
to those in the other classes. This suggests that the intensity pattern of fertile egg images
still has a relatively complex textural distribution.

In the Infertile class, the mean value was 116.7357, standard deviation was 25.2094, energy
value was 4.1353 × 10¹⁰, and entropy value was 5.1276. The standard deviation in this class
is the highest among the three classes, indicating greater pixel intensity variation in
infertile egg images. This condition suggests that infertile egg images have a more widely
distributed brightness variation, although their entropy value is slightly lower than that

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6876
of the Fertile class. Therefore, the standard deviation can be considered one of the
distinguishing features between the Infertile and other classes.

Meanwhile, the Highly Fertile class attained the highest mean value, 118.7481, with a
standard deviation of 21.2167, an energy value of 4.6269 × 10¹⁰, and an entropy value of
4.8324. The highest energy value indicates that images in the Highly Fertile class have
stronger dominant intensity in the LL sub-band. In contrast, the lowest entropy value
suggests that the image patterns in this class tend to be more structured or have a lower
information distribution than those in the other classes. The differences in mean, standard
deviation, energy, and entropy among the three classes indicate that the LL sub-band
provides distinguishable numerical patterns across egg fertility levels, supporting its use
as an input for the SVM classifier. These extracted statistical features play an important
role in the SVM classification process because they provide a numerical representation of
the intensity distribution and texture characteristics of each fertility category. The
differences in feature patterns among Highly Fertile, Fertile, and Infertile classes allow
the SVM classifier to construct a decision boundary for separating the three fertility levels.
Therefore, the Wavelet feature extraction stage contributes directly to the classification
performance by providing discriminative information rather than only reducing image
data into numerical values.

This study uses only the LL sub-band because it represents the approximation component
or the main information of the wavelet-transformed image. The LL sub-band stores low-
frequency information related to the global structure, intensity distribution, and main
shape of the egg object. In the context of egg fertility level classification, this global
information is considered more relevant because the differences among infertile, fertile,
and highly fertile classes are mainly reflected in the dominant intensity changes, bright–
dark regions, and overall visual patterns of the egg images. Therefore, the LL sub-band
was selected to obtain features that are more stable and less affected by noise or
irrelevant fine details.

In contrast, the LH, HL, and HH sub -bands mainly represent high-frequency detail
components, such as horizontal, vertical, and diagonal edges. Although these components
may provide additional information, the detail regions in the acquired egg images are
often affected by noise, shadows, eggshell texture, light reflections, and segmentation

Vol. 8, No. 5, October 2026

6877 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
artifacts. If the features from LH, HL, and HH are used directly, the extracted features may
capture not only fertility-related characteristics but also visual disturbances unrelated to
egg fertility levels. Thus, selecting the LL sub-band helps obtain a simpler, more stable,
and more focused feature representation based on the main intensity patterns of egg
images. This decision also reduces feature complexity before SVM classification,
particularly because this study uses statistical features such as mean, standard deviation,
energy, and entropy. However, the LH, HL, and HH sub-bands can still be explored in future
studies to evaluate the contribution of edge and detailed texture features in improving
egg fertility classification performance.

3.4. SVM Model Performance Evaluation
The evaluation of the SVM classification model was conducted in four stages: training
evaluation, testing evaluation, confusion matrix analysis, and computational efficiency
analysis. This sequence was used to evaluate the model learning capability, generalization
performance, classification behavior, and computational requirements. Egg fertility level
classification was performed using Support Vector Machine (SVM) based on the features
extracted from the Wavelet process. In this stage, the numerical features obtained from
feature extraction were used as input to distinguish egg images into three categories:
Highly Fertile, Fertile, and Infertile. SVM was selected because it can form an optimal
decision boundary between classes based on the given feature vectors. Thus, the model
can learn the characteristic patterns of egg images from the training data and use them
to predict the fertility class of new data.

Model performance was evaluated using two datasets: 1,200 training images and 300
testing images. The training data were used to examine the model’s ability to learn
patterns from the extracted Wavelet features, whereas the testing data were used to
measure the model’s generalization ability on images that were not used during the
training process. The classification results are presented in the form of confusion matrices
to show the relationship between the actual and predicted classes. In addition, model
performance was evaluated using accuracy, precision, recall, and the F1-score.

3.5. Training Evaluation
The first evaluation was performed using the training dataset consisting of 1,200 images.
This evaluation aims to measure the capability of the SVM model in learning fertility-

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6878
related patterns from the Wavelet features extracted from the egg images. The training
performance was analyzed using the confusion matrix and classification metrics, including
accuracy, precision, recall, and F1-score.

Table 5. Confusion Matrix Data Training
Actual / Predicted Highly Fertile Fertile Infertile Total
Highly Fertile 400 0 0 400
Fertile 52 291 57 400
Infertile 0 0 400 400
Total 452 291 457 1200

As shown in Table 5, the SVM model correctly classified 1,091 out of 1,200 training images.
The Highly Fertile and Infertile classes achieved perfect classification, with all 400 images
correctly predicted. However, several misclassifications occurred in the Fertile class, where
52 images were predicted as Highly Fertile and 57 images were predicted as Infertile. This
indicates that the Fertile category has more complex visual characteristics because it
represents an intermediate fertility condition between Highly Fertile and Infertile classes.

3.6. Testing Evaluation
After evaluating the training performance, the model was tested using an independent
testing dataset consisting of 300 images that were not involved during the training
process. This evaluation was conducted to measure the generalization capability of the
proposed SVM classifier on unseen egg images.

Table 6. Confusion Matrix Data Testing
Actual / Predicted Highly Fertile Fertile Infertile Total
Highly Fertile 69 31 0 100
Fertile 0 100 0 100
Infertile 0 0 100 100
Total 69 131 100 300

Vol. 8, No. 5, October 2026

6879 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
Based on Table 6, the SVM model correctly classified 269 out of 300 testing images. The
Fertile and Infertile classes achieved 100% correct classification, while the main error
occurred in the Highly Fertile class, where 31 images were classified as Fertil e. This
misclassification indicates that Highly Fertile and Fertile eggs may share similar intensity
distribution and texture characteristics in the LL Wavelet feature representation.

3.7. Confusion Matrix Analysis
The confusion matrix analysis shows that the dominant classification errors occurred
between the Highly Fertile and Fertile categories. Although both categories represent
fertile eggs, their visual characteristics during candling observation may have similar
intensity distributions and texture patterns. The LL Wavelet features used in this study
capture global intensity and structural information; however, some overlapping feature
distributions remain between these two fertility levels, making the classification boundary
more challenging for the SVM model.

In contrast, the Infertile category achieved complete classification in the testing dataset.
This indicates that infertile eggs have more distinctive visual characteristics compared
with fertile categories, allowing the extracted Wavelet features to provid e clearer
separation. Therefore, the confusion matrix analysis confirms that the proposed K-Means–
Wavelet–SVM framework is capable of distinguishing three fertility levels, while further
improvement may focus on reducing confusion between closely related fertile categories.

3.8. Computational Efficiency Analysis
In addition to classification performance, computational efficiency was evaluated to
determine the processing requirements of the proposed framework. Table 7 presents the
computation time and classification performance of the SVM model.

Table 7. Result SVM Model Performance Evaluation
Data Computation Time Accuracy Precision Recall F1-Score
Training 1.562 0.909 0.920 0.909 0.905
Testing 0.836 0.897 0.921 0.897 0.894

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6880
Based on Table 7, the SVM model achieved an accuracy of 0.909 on the training data and
0.897 on the testing data. These results show that the model produced good classification
performance on both the training and testing datasets. The difference in accura cy
between the training and testing data was relatively small (0.012), indicating that the model
did not show significant overfitting. In other words, the model was not only able to learn
patterns from the training data but also maintain its performance when applied to new
unseen data.

A precision value of 0.921 on the testing data indicates that the model predictions for egg
fertility classes had a high level of correctness. Meanwhile, a recall value of 0.897 shows
that the model was able to recognize most images from each fertility class. An F1-score
of 0.894 indicates a good balance between precision and recall, demonstrating that the
model was consistent in recognizing the three fertility categories.

The computational results also show that the proposed framework requires only 1.562
seconds for training and 0.836 seconds for testing, indicating efficient processing
capability. These results suggest that the proposed K-Means–Wavelet–SVM framework has
potential for practical implementation in automated egg fertility assessment.

3.9. Comparison with Previous Studies
A comparison with previous studies was conducted to demonstrate the position,
contribution, and advantages of the proposed method. This study integrates three main
stages–K-means segmentation, Wavelet feature extraction, and SVM classification–into a
structured workflow. This approach differs from several previous studies that used only
partial processing stages separately, such as focusing only on segmentation and
classification or using Wavelet features without applying segmentation as an initial step
to obtain the region of interest. By applying segmentation before feature extraction, the
analyzed images become more focused on regions relevant to egg fertility indicators,
making the extracted features more representative for classification. A comparison with
previous studies is presented in Table 8.

Vol. 8, No. 5, October 2026

6881 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
Table 8. Comparison with Previous Studies
Researcher Object Method Accuracy
Saha & Manickavasagan
(2021)
Egg fertility K-means + SVM 84%
Shi et al. (2022) Egg quality Wavelet + SVM/LDA/DT 78.5%–82.5%
Elibol et al. (2020) Egg classification DWT + SVM/LBNC
78.21%–
80.4%
This study Egg fertility
K-means + Wavelet +
SVM
89.4%

Based on Table 8, the proposed method achieved an accuracy of 89.4%, which is higher
than several previous studies using related approaches on egg fertility or egg quality
classification. Saha and Manickavasagan (2021) achieved an accuracy of 84% for egg
fertility classification using K-means and SVM. Meanwhile, Shi et al. (2022) obtained an
accuracy range of 78.5%–82.5% using Wavelet features with several classifiers, while Elibol
et al. (2020) achieved 78.21%–80.4% using DWT with SVM/LBNC classification methods. The
improved performance of the proposed K-Means–Wavelet–SVM pipeline can be attributed
to the sequential integration of image segmentation, feature extraction, and classification
processes. Unlike previous approaches that directly utilize the whole image or apply
feature extraction without prior region isolation, the proposed method first applies La-
based K-Means segmentation to reduce irrelevant background information. Furthermore,
the LL Wavelet sub-band provides a compact representation of global inte nsity
distribution and texture characteristics related to fertility development, while SVM
effectively separates the extracted feature patterns among the three fertility categories.
The advantage of the proposed method is the use of K-means as an initial stage to obtain
relevant image regions. Therefore, feature extraction is not performed on the entire raw
image but on a more focused region of interest. Wavelet Transform is then used to
generate more discriminative texture and intensity features, and the SVM performs the
classification based on these extracted features. Thus, the K -means–wavelet–SVM
approach not only improves the classification performance but also provides a more
systematic, objective, and measurable processing workflow for classifying egg fertility
levels. The improvement in classification performance can be attributed to the sequential
integration of segmentation, feature extraction, and classification stages in the proposed

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6882
pipeline. Unlike approaches that directly classify raw images or apply feature extraction
without prior region isolation, the proposed framework first applies La-based K-Means
segmentation to reduce irrelevant background information and obtain a more focused
egg region. Subsequently, LL Wavelet feature extraction provides compact statistical
representations of intensity and texture characteristics, while SVM classification
effectively separates the extracted feature patterns among the three fertility categories.

3.10. Discussion
The results of this study indicate that classifying egg fertility levels requires a structured
processing workflow, starting from segmentation, feature extraction, and classification. K-
means segmentation plays an important role in separating the region of interest in egg
images from the background, shadows, and other irrelevant image regions. This finding
indicates that segmentation is not only used to improve visual separation but also reduces
irrelevant variations before feature extraction. By focusing the analysis on the egg region,
the extracted features become more representative of fertility-related characteristics and
improve the reliability of the subsequent classification process.

Wavelet feature extraction contributes significantly by representing texture
characteristics, intensity patterns, and visual details of egg images in a multi-resolution
manner. The extracted statistical features, including mean, standard deviation, energy, and
entropy, provide quantitative representations of intensity distribution and texture
characteristics for each fertility category. These feature differences become important
inputs for the SVM classifier to construct decision boundaries among Highly Fertile, Fertile,
and Infertile classes.

In this study, the LL sub-band was selected because it provides a stable representation of
low-frequency information related to the overall intensity distribution and structural
characteristics of egg images. The selection of the LL sub-band represents a balance
between preserving relevant fertility information and reducing sensitivity to unnecessary
high-frequency variations. High-frequency components may contain detailed patterns
related to eggshell texture, illumination reflection, or acquisition noise rather than fertility
characteristics. Compared with the LH, HL, and HH sub-bands, the LL component is more
suitable for this study because egg fertility classification mainly depends on global
intensity changes and structural patterns observed during candling. In contrast, LH, HL,

Vol. 8, No. 5, October 2026

6883 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
and HH mainly represent high-frequency details, such as edges and local variations, which
may introduce less relevant information for fertility-level discrimination. Therefore, LL
features provide a more stable and compact feature representation that supports the
SVM classification process. In addition, the LL sub-band reduces the influence of high-
frequency variations associated with eggshell surface noise, uneven illumination during
candling, and reflection artifacts. By minimizing the contribution of these irrelevant
variations, LL features provide more consistent texture and intensity information related
to fertility characteristics, thereby improving the reliability of the classification process.

Compared with previous studies, the proposed approach provides an improvement
through the integration of K-means segmentation, Wavelet feature extraction, and SVM
classification into a systematic workflow. Previous research, such as fertile and infertile
egg classification on the fourth day of incubation, achieved an accuracy of 84%. The
improved performance of the proposed framework can be attributed to the use of La-
based K-means segmentation to reduce irrelevant image regions and LL Wavelet features
to capture fertility-related intensity and texture characteristics before classification.
Although the proposed framework achieved good classification performance, some
confusion remained between Highly Fertile and Fertile categories. This limitation is related
to the similarity of their visual patterns and overlapping feature distributions in the LL
Wavelet representation. These conditions make the SVM decision boundary more
challenging to separate the two closely related fertility levels.

Despite achieving promising performance, several limitations should be considered. First,
the dataset used in this study was obtained under relatively controlled image acquisition
conditions. Therefore, the performance of the proposed framework may vary wh en
applied to more complex industrial hatchery environments. Variations in illumination
conditions, camera devices, egg positioning, background conditions, and image acquisition
settings may affect the segmentation and feature extraction processes. Second, although
the dataset was balanced among the three fertility categories, its size is still limited
compared with the diversity of conditions found in real-world hatchery operations. Third,
this study focuses on image-based classification performance and does not yet evaluate
real-time processing capability in an automated industrial system.

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6884
From an industrial application perspective, the proposed Lab* -based K-Means
segmentation, LL Wavelet feature extraction, and SVM classification framework has
potential to be integrated into automated egg-candling conveyor systems for hatchery
management. Such implementation can utilize fixed cameras equipped with controlled
illumination to automatically capture egg images during the candling process. The
captured images can then be processed sequentially through image segmentation, feature
extraction, and classification stages to provide automatic fertility-level assessment. This
integration has the potential to reduce manual candling workload, improve inspection
consistency, and support more efficient decision-making in industrial hatchery operations.
For practical deployment, several technical considerations should be addressed, including
camera positioning, illumination consistency, synchronization between conveyor
movement and image acquisition, and processing speed requirements for real -time
inspection. Although the proposed SVM-based framework demonstrates efficient
computational performance, further validation using industrial-scale datasets and real-
time hardware platforms is required to evaluate its robustness under actual hatchery
conditions.

Future studies should further validate the proposed framework using larger datasets
collected under diverse hatchery conditions and evaluate its robustness against variations
in real-world image acquisition environments. In addition, further investigation involving
real-time implementation, optimization of processing speed, and comparison with
advanced segmentation and classification approaches can be conducted to improve the
generalization capability and practical applicability of the proposed framework.

4. CONCLUSION

This study successfully developed an image-based egg fertility level classification
framework by integrating K-means segmentation in the Lab* color space, low-low (LL)
sub-band Wavelet Transform for feature extraction, and Support Vector Machine (SVM)
for classification. Unlike previous studies that mainly focused on binary egg fertility
classification, this study classified eggs into three fertility levels: highly fertile, fertile, and
infertile. The proposed framework was evaluated using a dataset consisting of 1,500 egg
images, with 1,200 images used for training and 300 images used for testing. The
experimental results demonstrated that the SVM classifier achieved 90.9% accuracy, 92.0%

Vol. 8, No. 5, October 2026

6885 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
precision, 90.9% recall, and 90.5% F1-score on the training dataset, while achieving 89.7%
accuracy, 92.1% precision, 89.7% recall, and 89.4% F1-score on the testing dataset. The small
performance difference between training and testing results indicates that the proposed
framework provides good generalization capability with limited overfitting. The main
contribution of this study lies in extending conventional binary egg fertility detection into
a multilevel fertility classification framework by combining La-based K-means
segmentation, LL Wavelet feature representation, and SVM classification. From a practical
perspective, the proposed framework has the potential to support automated hatchery
management by enabling more objective fertility screening, reducing manual candling
workload, and improving incubation efficiency through earlier identification of fertility
levels. However, this study was conducted using egg images acquired under controlled
conditions; therefore, further validation using more diverse hatchery environments,
including variations in illumination, camera distance, egg orientation, and acquisition
conditions, is required to further assess the robustness and practical applicability of the
proposed framework.

REFERENCES

[1] X. Yang, R. B. Bist, S. Subedi, and L. Chai, “A Computer Vision-Based Automatic System
for Egg Grading and Defect Detection,” Animals, vol. 13, no. 14, 2023, doi:
10.3390/ani13142354.
[2] R. L. M. Bambilla, M. J. B. Lee, and F. L. Valiente, “Freshness Condition Detection, Size
Classification with Weight Prediction and Sortation of Philippine Egg Produce Using
YOLOv8 and OpenCV,” in 2025 17th International Conference on Computer and
Automation Engineering, ICCAE 2025, School of Electrical, Electronics and Computer
Engineering, Mapúa University, Manila, Philippines: Institute of Electrical and
Electronics Engineers Inc., 2025, pp. 158–164. doi: 10.1109/ICCAE64891.2025.10980499.
[3] H. A. Muslim, H. Oktavianto, R. T. Widodo, E. Purwantini, and W. T. Sesulihatien,
“Implementation of Chicken Eggs Fertility Detection Device Using SSD MobileNet-
V2 FPNLite,” 2024 Int. Electron. Symp. Shap. Futur. Soc. 5.0 Beyond, IES 2024 -
Proceeding, pp. 171–177, 2024, doi: 10.1109/IES63037.2024.10665784.
[4] D. Sun, H. Wei, Y. Luo, W. Li, and Y. Yu, “Real-Time Detection Method for Cracked Eggs
Based on Improved YOLOv8 and Dual Verification with Triple Classification
Granularity,” IEEE Access, 2026, doi: 10.1109/ACCESS.2026.3686452.

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6886
[5] S. L. de Castro Júnior, A. E. C. M. Cândido, A. C. de S. Silva, and I. J. O. da Silva,
“Translucency of chicken eggs: proposal of an automated system for analysis and
classification,” Poult. Sci., vol. 104, no. 10, 2025, doi: 10.1016/j.psj.2025.105612.
[6] T. D. Pham, “Classification of Motor-Imagery Tasks Using a Large EEG Dataset by
Fusing Classifiers Learning on Wavelet-Scattering Features,” IEEE Trans. Neural Syst.
Rehabil. Eng., vol. 31, pp. 1097–1107, 2023, doi: 10.1109/TNSRE.2023.3241241.
[7] P. Zeng et al., “High-Accuracy Classification of Fertilized Eggs Using Mamba-Based
Model,” in 2025 6th International Conference on Artificial Intelligence and
Electromechanical Automation, AIEA 2025, School of Physics and Information
Engineering, Guangdong University of Education, Guangzhou, 510303, China:
Institute of Electrical and Electronics Engineers Inc., 2025, pp. 92 –96. doi:
10.1109/AIEA66061.2025.11159668.
[8] P. G. Gülhan and G. Özmen, “Follicle Detection for Polycystic Ovary Syndrome by
Using Image Processing Methods,” Int. J. Appl. Math. Electron. Comput., vol. 8, no. 4,
pp. 203–208, 2020, doi: 10.18100/ijamec.803400.
[9] C. Shi et al., “Nondestructive Detection of Microcracks in Poultry Eggs Based on the
Electrical Characteristics Model,” Agric., vol. 12, no. 8, p. 1137, 2022, doi:
10.3390/agriculture12081137.
[10] S. Anantharajan and S. Gunasekaran, “Automated brain tumor detection and
classification using weighted fuzzy clustering algorithm, deep auto encoder with
barnacle mating algorithm and random forest classifier techniques,” Int. J. Imaging
Syst. Technol., vol. 31, no. 4, pp. 1970–1988, 2021, doi: 10.1002/ima.22582.
[11] I. K. G. D. Putra, N. P. A. O. Wiastini, K. S. Wibawa, and I. M. S. Putra, “Identification of
Skin Disease Using K-Means Clustering, Discrete Wavelet Transform, Color Moments
and Support Vector Machine,” Int. J. Mach. Learn. Comput., vol. 10, no. 4, pp. 542–548,
2020, doi: 10.18178/ijmlc.2020.10.4.970.
[12] D. M. Toufiq, A. M. Sagheer, and H. Veisi, “Brain tumor identification with a hybrid
feature extraction method based on discrete wavelet transform and principle
component analysis,” Bull. Electr. Eng. Informatics, vol. 10, no. 5, pp. 2588–2597, 2021,
doi: 10.11591/eei.v10i5.3013.
[13] F. A. Jibon et al., “Cancerous and Non-Cancerous Brain MRI Classification Method
Based on Convolutional Neural Network and Log-Polar Transformation,” Healthc.,
vol. 10, no. 9, p. 1801, 2022, doi: 10.3390/healthcare10091801.

Vol. 8, No. 5, October 2026

6887 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
[14] S. K. Prabhakar and D. O. Won, “Performance comparison of bio -inspired and
learning-based clustering analysis with machine learning techniques for
classification of EEG signals,” Front. Artif. Intell., vol. 6, 2023, doi:
10.3389/frai.2023.1156269.
[15] D. Saha and A. Manickavasagan, “Machine learning techniques for analysis of
hyperspectral images to determine quality of food products: A review,” Curr. Res.
Food Sci., vol. 4, pp. 28–44, 2021, doi: 10.1016/j.crfs.2021.01.002.
[16] G. Elibol, M. Yumurtaci, S. Ergi̇n, and İ. Yabanova, “Classification of Dynamic Egg
Weights Using Feature Extraction Methods,” Eskişehir Tech. Univ. J. Sci. Technol. a -
Appl. Sci. Eng., vol. 21, no. 4, pp. 499–513, 2020, doi: 10.18038/estubtda.658077.
[17] A. O. Adegbenjo, L. Liu, and M. O. Ngadi, “Non-destructive assessment of chicken egg
fertility,” Sensors (Switzerland), vol. 20, no. 19, pp. 1–23, 2020, doi: 10.3390/s20195546.
[18] D. Kalpana and K. S. Reddy, “SVM for Classification of Brain MRI with K-Means for
Detection,” Interantional J. Sci. Res. Eng. Manag., vol. 07, no. 11, pp. 1–11, 2023, doi:
10.55041/ijsrem26641.
[19] G. Yuan, Y. Liu, W. Huang, and B. Hu, “Differentiating Grade in Breast Invasive Ductal
Carcinoma Using Texture Analysis of MRI,” Comput. Math. Methods Med., vol. 2020,
pp. 1–14, 2020, doi: 10.1155/2020/6913418.
[20] A. Nanda, R. C. Barik, and S. Bakshi, “SSO-RBNN driven brain tumor classification with
Saliency-K-means segmentation technique,” Biomed. Signal Process. Control, vol. 81,
2023, doi: 10.1016/j.bspc.2022.104356.
[21] S. Fooladi, H. Farsi, and S. Mohamadzadeh, “Segmenting the Lesion Area of Brain
Tumor using Convolutional Neural Networks and Fuzzy K-Means Clustering,” Int. J.
Eng. Trans. B Appl., vol. 36, no. 8, pp. 1556–1568, 2023, doi: 10.5829/IJE.2023.36.08B.15.
[22] O. S. Faragallah, H. M. El-Hoseny, and H. S. El-sayed, “Efficient brain tumor
segmentation using OTSU and K-means clustering in homomorphic transform,”
Biomed. Signal Process. Control, vol. 84, 2023, doi: 10.1016/j.bspc.2023.104712.
[23] A. Saday and İ. A. Ozkan, “Classification of Epileptic EEG Signals Using DWT-Based
Feature Extraction and Machine Learning Methods,” Int. J. Appl. Math. Electron.
Comput., vol. 9, no. 4, pp. 122–129, 2021, doi: 10.18100/ijamec.988691.
[24] I. Zine-Dine, J. Riffi, K. El Fazazy, I. El Batteoui, M. A. Mahraz, and H. Tairi, “A review:
Machine Learning Techniques of Brain Tumor Classification and Segmentation,”
Mach. Graph. Vis., vol. 34, no. 3, pp. 31–55, 2025, doi: 10.22630/MGV.2025.34.3.2.

Vol. 8, No. 5, October 2026

Sriyadi, Jufriadif Na'am, et al | 6888
[25] A. Tayeh Swadi and F. Sabar Miften, “Detection of attention deficit hyperactivity
disorder based on EEG signals using Least Square Support Vector Machine (LS-
SVM),” Comput. Methods Biomech. Biomed. Eng. Imaging Vis., vol. 11, no. 6, pp. 2495–
2507, 2023, doi: 10.1080/21681163.2023.2243347.
[26] K. P. Ajitha Gladis, A. Ahilan, N. Muthukumaran, and L. Jenifer, “ECG Arrhythmia
Measurement and Classification for Portable Monitoring,” Meas. Sci. Rev., vol. 24, no.
4, pp. 118–128, 2024, doi: 10.2478/msr-2024-0017.
[27] P. Yan, Y. Zhang, S. Hou, W. Zhang, and X. Yang, “Noise reduction method for blasting
vibration signals based on improved EMD -wavelet packet,” Zhendong yu
Chongji/Journal Vib. Shock, vol. 43, no. 11, pp. 264 -271and287, 2024, doi:
10.13465/j.cnki.jvs.2024.11.029.
[28] H. Yang, Y. Zhai, J. Liang, D. Guo, X. Liu, and R. Zhang, “Algorithm for Extracting Wear
Characteristics of Piston Pins Based on Orthogonal Empirical Mode Decomposition,”
Nongye Jixie Xuebao/Transactions Chinese Soc. Agric. Mach., vol. 55, pp. 412–419,
2024, doi: 10.6041/j.issn.1000-1298.2024.S1.044.
[29] K. K. Çevik, H. E. Koçer, and M. Boğa, “Deep Learning Based Egg Fertility Detection,”
Vet. Sci., vol. 9, no. 10, 2022, doi: 10.3390/vetsci9100574.
[30] J. L. Dioses, R. P. Medina, A. C. Fajardo, and A. A. Hernandez, “Performance of
Classification Models in Japanese Quail Egg Sexing,” in Proceeding - 2021 IEEE 17th
International Colloquium on Signal Processing and Its Applications, CSPA 2021 ,
Graduate School, Technological Institute of the Philippines, Quezon City, Philippines:
Institute of Electrical and Electronics Engineers Inc., 2021, pp. 29 –34. doi:
10.1109/CSPA52141.2021.9377275.
[31] Y. F. Wei, P. Yang, Z. Y. Yang, P. Wang, and X. W. Wang, “Fault Detection of Flexible DC
Grid Based on Empirical Wavelet Transform and WOA-CNN,” J. Electr. Eng. Technol.,
vol. 20, no. 1, pp. 217–229, 2025, doi: 10.1007/s42835-024-02038-9.
[32] M. H. Saleem, J. Potgieter, and K. M. Arif, “Automation in Agriculture by Machine and
Deep Learning Techniques: A Review of Recent Developments,” Precision
Agriculture, vol. 22, no. 6. pp. 2053–2091, 2021. doi: 10.1007/s11119-021-09806-x.
[33] S. Saifullah, R. Drezewski, A. Khaliduzzaman, L. K. Tolentino, and R. Ilyos, “K-Means
Segmentation Based-on Lab Color Space for Embryo Detection in Incubated Egg,”
J. Ilm. Tek. Elektro Komput. dan Inform. , vol. 8, no. 2, p. 175, 2022, doi:
10.26555/jiteki.v8i2.23724.

Vol. 8, No. 5, October 2026

6889 | A Computer Vision-Based Framework for Multilevel Egg Fertility Classification …..
[34] V. Kiruthika and M. M. Ramya, “Automatic segmentation of ovarian follicle using K-
means clustering,” Proc. - 2014 5th Int. Conf. Signal Image Process. ICSIP 2014, pp.
137–141, 2014, doi: 10.1109/ICSIP.2014.27.