2022 8th International Engineering, Sciences and Technology Conference (IESTEC)

Fruit Detection and Classification Using Computer

Vision Techniques
sşĐƚŽƌĄƌĂƚĞ ĂŶŝůŽĄĐĞƌĞƐ
&ĂĐƵůƚĂĚĚĞ/ŶŐĞŶŝĞƌşĂůĠĐƚƌŝĐĂ &ĂĐƵůƚĂĚĚĞ/ŶŐĞŶŝĞƌşĂůĠĐƚƌŝĐĂ
hŶŝǀĞƌƐŝĚĂĚdĞĐŶŽůſŐŝĐĂĚĞWĂŶĂŵĄ hŶŝǀĞƌƐŝĚĂĚdĞĐŶŽůſŐŝĐĂĚĞWĂŶĂŵĄ
WĂŶĂŵĄ WĂŶĂŵĄ
ǀŝĐƚŽƌ͘njĂƌĂƚĞΛƵƚƉ͘ĂĐ͘ƉĂ ĚĂŶŝůŽ͘ĐĂĐĞƌĞƐΛƵƚƉ͘ĂĐ͘ƉĂ
2022 8th International Engineering, Sciences and Technology Conference (IESTEC) | 978-1-6654-9843-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/IESTEC54539.2022.00110

Abstract—Fruit detection and classification are very x Electric noses: These examine the chemical
important processes that can lead to automation improvements composition of the volatiles that are produced by the
in agricultural and logistical tasks and therefore lead to a better
industrialization of this sector. It makes these tasks more
fruit [3]. These can also detect Ethylene level, which
efficient since they usually need a lot of human labor. In this is a key indicator of fruit ripening [4].
paper an algorithm is developed for the detection and x Soft Tactile sensors: These sensors work by applying
classification of fruits using image processing and computer a very small force on the fruit, as to not destroy it,
vision techniques. This will be made using a coding software and sensing the fruits response [5].
with no external libraries, this way the working mechanisms of
the process can be better understood, making it a good x Computer Vision: These use the visual information
educational resource for students that could be interested into of the fruit, to extract characteristics, and use an
learning computer vision. intelligent system to make decisions. As information
science is rapidly growing, computer vision-based
Keywords—classifier, object detection, computer vision
pattern recognition and image processing are
I. INTRODUCTION matured techniques for safety and quality analysis of
Fruit detection and classification are very important several agricultural applications [6].
processes that can lead to automation improvements in Another important aspect of fruit recognition,
agricultural and logistical tasks and therefore lead to a better classification and quality determination is the method
industrialization of this sector. It makes these tasks more implemented to process the information and characteristics
efficient since they usually need a lot of human labor. One of observed by the sensors. In 2014, a method was proposed and
the most important aspects of fruit recognition and tested consisting of a fruit classification algorithm which used
classification is the sensing method. These methods can be merely a digital camera, this algorithm used feedforward
divided in two distinguishing categories: invasive and neural network techniques [7]. A paper in 2012, just two years
noninvasive technologies. These sensing mechanisms also prior, proposed a method that used a support vector machine
make it possible to determine the perceived and actual quality (SVM) classifier [8]. In 2015, a research paper made a review
of the fruit. Some of the implemented mechanisms are: of the methods available at the time for the automatic
1. Invasive: classification and grading of quality of fruit using computer
vision using mainly a camera as the sensing component [9].
x Sweetness sensors: This method involves Another paper with the same goal of reviewing the methods
quantifying the sugar content of the fruit, usually available at the time was also published in 2017, stating that
measured in brix. In some cases, they rely on one of the main quality features of fruits is its appearance,
dissolving solid parts of the fruit in samples since it influences their market value, the preference and
effectively destroying the fruit. In other cases, choices of the consumer, and can also be an indicator of the
microextraction methods can be implemented, internal quality of the fruit. The scope of the sensing
mechanisms was a bit broader in this paper, since it used other
however, these are not commercially available [1]. kind of sensor aside from a digital camera, like weighting
x PH and acidity: This method involves quantifying the machines, firmness tester, and aroma sensor [10]. Another
concentration of hydrogen ions in a solution made paper like this was also published in 2021, where even more
from the fruit [1]. A measurement can also be made image acquisition techniques were mentioned, like ultrasound,
directly from the fruit, although piercing the fruit infrared, and tomographic imaging. In this work, early defect
with the sensing device is required in this case. detection was mentioned as an important task, but also a very
challenging one due to the huge variation of defect types [6].
2. Noninvasive: In this paper an algorithm is developed to detect and
x Weight sensors: These include simple scales which classify fruits using computer vision techniques. The goal of
don’t damage the fruit. this paper is to introduce an algorithm that classifies fruit from
a known subset of fruit types in a fast and computationally
x Moisture sensors: These include magnetic resonance inexpensive way. In order to maintain it computationally
imaging, near infrared spectroscopy and hyper- inexpensive, a method consisting of only extracting the main
spectral imaging [2]. color and circularity of the fruit was proposed. By keeping the
computation cost low, this method is expected to be scalable

978-1-6654-9843-2/22/$31.00 ©2022 IEEE 665

DOI 10.1109/IESTEC54539.2022.00110
Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 so that it can classify from a bigger subset of fruits and in some the threshold. If it is above, then it is considered a background
future for it to be able to detect fruit quality. pixel (since it is assumed that all the fruit images have a white
background). And if it is below that threshold value, then it is
This algorithm will be implemented using the SciLab considered a fruit pixel.
Software and without the use of any external libraries since
the goal is to make this implementation as educational and Every pixel in the image have now been classified as either
reproducible as possible. The synopsis of the process is as a background pixel or fruit pixel. The intensity value of all the
follows: background pixels is now set as 0 so the area that has been
classified as background can be visualized and proceed to
1. Segmentation of the fruit: This eliminates insignificant erase every column and row that is only composed of
information and brings focus on the important background pixels. This is to get rid of all the irrelevant
information, in this case being the fruit to be analyzed information and basically “zoom” into the fruit.
and classified.
B. Resolution Downscaling
2. Resolution downscaling: This is to make the process
more manageable, since it reduces the amount of To reduce the image size a “BOX” filter was used, this is
information that must be processed. a low pass separable filter with a box kernel, in which all the
coefficients of this kernel have the same value (usually 1). A
3. Color extraction: This is a required descriptor to box filter is a fundamental filter in image processing [11]. Its
classify the fruits. name, Box Kernel, is derived from the fact that a constant
4. Circularity extraction: This is a required descriptor to kernel resembles a box when visualized in 3-D, several box
classify the fruits. filter techniques have been developed, as seen in [12], in this
paper this filter will be used to downscale the fruit image. A
5. Classifier: This process identifies the fruit that is being “Box” filter is a matrix of ones, with a normalization constant,
analyzed based on the extracted descriptors (color and which value is 1 divided by the number of coefficients, 1/mn.
circularity), by comparing them with a given dataset in This normalization applied to all the low pass nuclei has two
a probabilistic manner. purposes. First, the average intensity value of a constant
This paper begins by describing the materials and intensity area would be equal to that same intensity. Secondly,
methods, a description of the concepts and theories behind of normalizing the kernel this way avoids the introduction of a
each necessary process for the detection and classification of bias during filtering; that is, the sum of the intensity in the
the fruit is given. Then the results and observations are shown. original image and the filtered image will be the same. Each
In Fig. 1 a flowchart indicating the main processes and their pixel value in the output image is the arithmetic mean of the
relation is shown. neighbor pixels of the reference pixel. The resulting kernel can
be observed in (1):

ͳ ͳ ‫ͳ ڮ‬
‫ܯ‬ൌ ή ൥‫ڭ‬ ‫ڭ ڰ‬൩ ;ϭͿ
݉݊
ͳ ‫ͳ ڮ‬

Now that the kernel is constructed, the image can now be

divided in square boxes of pixels, the total amount of boxes
will be the goal number of pixels. The kernel is now applied
to every box, and the resulting intensity will be mapped as the
intensity of that pixel in the resulting image.
C. Color Extraction
The method used to extract the color as to optimize the
“precision/computational complexity” ratio in this case was a
K-Means clustering method, this method has been
implemented before for fast segmentation as seen in [13]. The
first step is to map every single pixel color into a point in a 3D
color space, where the three dimensions are the red, green, and
Fig. 1 Fruit detection and classification algorithm flowchart. blue intensity values.

II. PROPOSED METHOD Now, every one of these points are sorted into one of K
clusters. Since one of the goals is to optimize the
Having established a process roadmap for this goal, this “precision/computational complexity”, they will be grouped
section focuses on explaining more in depth the concepts and into 3 clusters. Since every fruit image has a main fruit color,
theories that are needed to make each step of this flowchart a color for the secondary characteristics of the fruit, and a
possible. background color (white in this case). The sorting takes place
A. Segmentation in a way so that the mean of the distances between every point
and the center of the corresponding cluster is minimum.
The process starts by turning the image into grayscale, the
goal of this is to make the distinction between fruit and The algorithm to be used works in the following way:
background easier. Now, a threshold value can be chosen and
1. First, the position of the center of the 3 clusters will be
check for each pixel if the intensity value is above or below
selected at random.

666

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 2. Each point will check which cluster center is closest
and will now be classified as part of that cluster.
3. The position of the center of the 3 clusters will now be
calculated using the positions of the points that are part
of that cluster.
4. If any cluster has no points, then the position of its
center will be located at random again.
5. Repeat steps 2 through 4 until convergence is achieved.
This algorithm will converge to a local minimum, the
global minimum is not guaranteed. The results of the
convergence will depend on the first random values that were
used. Since 3 clusters were used for this algorithm, and since
most of our images have 3 main colors, the local minimum
achieved will be sufficient for our purposes. Now to
demonstrate, the image of the pear that is 50 pixels by 39
pixels will be used as an example, in Fig. 2 the 3 main clusters
that were found can be seen and in Fig. 3 their center and how
they converged can be seen.

Fig. 3 The clusters means and how they converged.

The locations of the center of the 3 cluster have been

obtained, representing the 3 main colors of the image. The one
with the highest number of points in the cluster will now be
extracted, except if the intensity is too high, because that
would mean the white color of the white background was
extracted, in this case this cluster will be ignored and extract
the center of the next one.
The fruit color that was extracted is a 3-dimensional
descriptor, and therefore, if the circularity was added, the
feature space would be 4-dimensional. Since one goal is to
optimize the “precision/computational complexity” a shift is
needed to change the reference system to a more intuitive one,
closer to how humans perceive color. The HSV reference
system is now introduced.
The HSV reference system has 3 dimensions as well, those
being Hue, Saturation and Value (or brightness). But it is
important to point out that the saturation of the fruit could have
different factors affecting it, so it could be sub optimal for this
paper purposes. The same could be said for the brightness of
the color.
However, hue is a characteristic that human beings can
perceive instantly as soon as a classification is being made.
So, that intuition is given to this algorithm as well.
That means a descriptor for color that has only one
dimension has been obtained, hue.
D. Circularity Extraction
The method used to extract the circularity will start by
Fig. 2 The 3 main clusters that are found in the pear image. using the following equation:

667

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
   First, a set of already classified images of fruits is needed,
Ͷߨ‫ܣ‬ then the features of this images will be extracted, those being
‫ܥ‬ൌ ଶ ;ϮͿ their main color and circularity.
ܲ
 Now, they will be modeled as random variables and the
 mean and standard deviation for all the features of each fruit
Where C is the dimensionless circularity of the fruit, P is will be computed. This can be mapped into a 2-D plane in
the perimeter of the fruit in pixels, and A is the area of the fruit which one axes is the color and the other axes the circularity.
in pixels squared. A point for every single classified image can be mapped here.
If a point without classification is positioned into this map, the
Now, to obtain these parameters the resized image will be means, and their standard deviation of the classified set of
gray scaled and binarized using an intensity threshold. fruits can be used to classify that point into a specific fruit.
Once the image is separated into fruit pixels and To simplify the naïve Bayes method, a Manhattan distance
background pixels, by counting the amount of fruit pixels that between the unclassified point and the means of the classified
there are, the fruit area. sets will be used. This Manhattan distance will be normalized
to be in the units of standard deviations for each specific set.
To obtain the perimeter a Moore boundary tracing
Then the distances between the unclassified point and all the
algorithm will be implemented, these algorithm uses the
means of the classified set will be compared and the set with
binarized image to trace the boundary of all the fruit pixels,
the shortest Manhattan distance will be chosen as the set this
while doing so, a Freeman Chain code algorithm can be used
point will be classified into. the new image is now classified
to store the direction of the movements being made while
as a fruit.
tracing the boundary. By summing up exactly how many times
the trace moved horizontally and vertically and how many III. ALGORITHM DESCRIPTION AND
times it moved diagonally, if it now multiplies every IMPLEMENTATION
horizontal and vertical movement by one, and all the diagonal
movements by the square root of two and add all these lengths Now, equipped with a better understanding of the parts of
it will have as a result the perimeter of the fruit. Having the process and how they all relate, an executable computation
computed the area and the perimeter a computation of the for this procedure must be implemented, this will be the
circularity can now be made. This Moore boundary tracing algorithm. The algorithm must be designed in a way that
process can be visualized with an example in Fig. 4. maximizes the accuracy of the classification while minimizing
the computation power and computation time. This
implementation will be made using the SciLab software
without the use of external libraries, for educational and
clarity purposes.
A. Algorithm Description
Since the goal is to be able to make this classification
without the use of external libraries, a new library will be
made from scratch. This library will contain, in an organized
manner, all the necessary processes and algorithms that are
needed to extract the features of the fruit in the image, and to
Fig. 4 Moore boundary tracing algorithm, paired with a freeman
classify them accordingly.
chain code, to compute the perimeter of the pixelated image. x grayFromRGB: This function’s input is a colored
image, and its output is the same image converted into
Modified versions of Moore algorithms have been grayscale.
implemented recently as seen on [14].
x zeroFill: This function’s input is a grayscale image,
E. Classifier and Its output is the same image with an extra layer of
A simplified version of the naive bayes classifier will be zeroes.
implemented. The naive bayes classifier is a probability-based
classifier that is based on applying Bayes’ theorem with naïve x recortarAbove: this function’s inputs are a colored
independence assumptions between features [15]. such that: image and an integer between 0 and 255; this integer
will be the threshold, and its output is the same image
  but zoomed in without considering the background.
ܲሺ‫ܤ‬ȁ‫ܣ‬ሻܲሺ‫ܣ‬ሻ The way it works is that it first converts it into
ܲሺ‫ܣ‬ȁ‫ܤ‬ሻ ൌ  grayscale, then completely removes any row or column
ܲሺ‫ܤ‬ሻ ;ϯͿ that is entirely made up of intensities above the
 threshold. Then converts the smaller image back into a
 colored one.
In this example A is defined as the set of a fruit, and B as x scalingTo: this function’s input are a colored image
the set of descriptors (color or circularity) that could be and an integer; this integer will be the goal resolution.
defined by B=CģF where C is the color set and F is the The way it works is that it lowers the resolution of the
circularity set. input image, without changing its scale, until the
smallest axis is the closest possible integer value to the
goal resolution. If the goal resolution is larger than the

668

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 smallest axis, then this function does nothing. Its circularity of the fruit. The output of this function is the
output will be the lower resolution image, close to the dimensionless circularity of the fruit.
goal resolution.
B. Algorithm Implementation
x kMeansColorsAbove: This function’s input are a The first step is to choose the types of fruits that the
colored image and an integer; this integer will be the program will be able to categorize, for this implementation 7
threshold. Its function is to identify the 3 main different fruits were chosen, which were:
distinguishable colors in the image and to extract the
most abundant one. If the color is converted into 1. Apple
grayscale and its intensity value is above the threshold 2. Pear
then that means it is the background color, and the
color that will be extracted will be the second most 3. Orange
abundant one. The color obtained is in RGB reference 4. Banana
format and will be converted into HSV and the hue will
be extracted. The output of this function is the hue in 5. Green Lemon
degrees.
6. Yellow Lemon
x circularityAbove: this function’s input are a colored
7. Pineapple
image and an integer; this integer will be the threshold.
First it converts the image into grayscale and then Subsequently, 10 images of each fruit were downloaded
binarizes it, all the pixels with an intensity value above from the internet with the intention of obtaining the circularity
the threshold will be considered background pixels and and hue distribution for each one. To obtain this circularity
the ones with intensity values below the threshold will and hue distribution, each fruit image went through a
be considered fruit pixels. With this binarized image of preprocessing that consisted in cutting out the white
the fruit it now computes the perimeter and area of the background and downscaling the resolution as close as
fruit with the algorithms mentioned. Its purpose is to possible to 64 pixels squared without distorting its aspect
use the perimeter and area values to obtain the ratio. Then the circularity and hue are extracted. This process
can be visualized in Fig. 5.
&ƌƵŝƚ ^ĞŐŵĞŶƚĂƚŝŽŶ ŽǁŶƐĐĂůŝŶŐ ŝƌĐƵůĂƌŝƚLJ ŽůŽƌ

Fig. 5 Feature extraction process visualization

669

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 IV. ACHIEVED RESULTS
Once the implementation and debugging are carried out, the
results of this algorithm with the given implementation
parameters can be observed.
A. Feature Space
In Fig. 6 the hue and circularity distribution for each of the
7 different fruits can be observed, the X-axis being the
circularity, and the Y-axis being the hue.

Fig. 8 Classification range for all fruits.

In Fig. 9 the classification range of fruits when neither the

pineapples nor the yellow lemons are taken into consideration
can be observed:

Fig. 6 Hue and circularity distribution for all the 7 fruits. The X-
axis represents the circularity which is dimensionless. The Y-axis
represents the hue in degrees.

The red dots are the apples, the green dots are the pears,
the pink dots are the oranges, the yellow dots are the bananas,
the cyan dots are the pineapples, the black dots are the yellow
lemons, and the blue dots are the limes.
Next, in Fig. 7 the same distributions can be seen but
without taking the pineapples and yellow lemons into account.

Fig. 9 Classification range for all fruits, except for pineapples and
yellow lemons.

B. Classification
In order to test the program, the goal to classify 10 images
of each fruit that are not part of the training dataset of the
program was proposed. 5 images for each fruit are
downloaded from the internet, and the other 5 images were
taken with a cellphone camera and the fruit was placed in front
of a white cardboard. In some of these images the brightness
and contrast were adjusted to further accentuate the
background.
Classifying all the 70 images took 1 hour and 35 minutes.
In Table 1 the results of the classification for each of the 10
images of each fruit can be seen. And in Table 2 the confusion
matrix for these results can be seen.
Fig. 7 Hue and circularity distribution for all the 7 fruits, except The confusion matrix, also known as the error matrix, is
for the pineapple and yellow lemon. The X-axis represents the
the table used for visualizing the performance of an algorithm
circularity which is dimensionless. The Y-axis represents the hue in
and for calculation of various evaluation metrics. It provides
degrees.
an insight about the types of errors that are made by the
In Fig. 8 the classification range for all fruits can be classifier [16].
observed:

670

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 Table 1. Classification Metrics In Fig. 10 it can be observed that the body of the yellow
lemon is considered as a background pixel due to the lighting
DĞƚƌŝĐƐ at the time the picture was taken.
Input T T F F Sensitivit Specificit F1-
Fruit P N P N y y Scor
e
Apple 9 60 Ϭ 1 0.9 1 0.94
7
Pear 10 59 ϭ 0 1 0.983 0.95
2 Fig. 10 Inaccuracy caused by the lighting of the yellow lemon.

Orang 10 56 ϰ 0 1 0.933 0.83 In Fig. 11 it can be observed that the image has been split
e 3 into more than one region, this causes the perimeter to be
Banan 9 59 ϭ 1 0.9 0.983 0.9 calculated only on the top region, due to the way the circularity
a algorithm works, this causes the circularity to have a wrong
value, sometimes above one, which is impossible.
G. 10 60 Ϭ 0 1 1 1
Lemon
Y. 7 60 Ϭ 3 0.7 1 0.82
Lemon 4
Pineappl 9 60 Ϭ 1 0.9 1 0.94
e
7

Table 2. Confusion Matrix

Fig. 11 Inaccuracy caused by the bright color of the banana.
ZĞƐƵůƚƐ
/ŶƉƵƚ ůĂƐƐŝĨŝĐĂƚŝŽŶ
'͘ z͘
In Fig. 12 the same case can be observed for the yellow
YƵĂŶƚŝƚLJ &ƌƵŝƚ ƉƉůĞ WĞĂƌ KƌĂŶŐĞ ĂŶĂŶĂ >ĞŵŽŶ >ĞŵŽŶ
WŝŶĞĂƉƉůĞ
lemon, however, in this case the repercussions are worse,
ϭϬ ƉƉůĞ ϵ Ϭ ϭ Ϭ Ϭ Ϭ Ϭ
ϭϬ WĞĂƌ Ϭ ϭϬ Ϭ Ϭ Ϭ Ϭ Ϭ given that the first region consists of only one pixel, this means
ϭϬ KƌĂŶŐĞ Ϭ Ϭ ϭϬ Ϭ Ϭ Ϭ Ϭ the Moore boundary tracing algorithm gets stuck with no more
ϭϬ ĂŶĂŶĂ Ϭ ϭ Ϭ ϵ Ϭ Ϭ Ϭ
ϭϬ '͘>ĞŵŽŶ Ϭ Ϭ Ϭ Ϭ ϭϬ Ϭ Ϭ
pixels to trace, this causes the program to get stuck and
ϭϬ z͘>ĞŵŽŶ Ϭ Ϭ ϯ Ϭ Ϭ ϳ Ϭ completely freeze.
ϭϬ WŝŶĞĂƉƉůĞ Ϭ Ϭ Ϭ ϭ Ϭ Ϭ ϵ


V. CONCLUSION
It can be observed in the feature space that the fruits have
visible continuous regions in which they would classify an
unknown point, this happens for both cases, when the
pineapple and yellow lemon are considered, and when they
Fig. 12 Inaccuracy caused by the lighting in the yellow lemon.
are not.
Another observation is that the hue of the fruit is the most REFERENCES
determining factor when the fruit is classified, this could be [1] D. Wilson, "Chemical Sensors for Farm-to-Table Monitoring of
due to a lack of precision in the program when extracting the Fruit Quality", Sensors, vol. 21, no. 5, p. 1634, 2021, pp.1-29.
circularity, and because there are few cases where there are [2] A. Ren et al., "Machine Learning Driven Approach Towards the
fruits with identical colors but contrasting circularities. Quality Assessment of Fresh Fruits Using Non-Invasive
However, this can be observed in the case for bananas and Sensing", IEEE Sensors Journal, vol. 20, no. 4, 2020, pp. 2075-
2083.
yellow lemons, where both coexist in a very narrow hue range,
[3] M. Baietto and A. Wilson, "Electronic-Nose Applications for
but are distinguishable mostly by their circularity. Fruit Identification, Ripeness and Quality Grading", Sensors, vol.
15, no. 1, 2015, pp. 899-931.
One of the most time intensive parts of the project was the
debugging process and making sure it ran in an optimal and [4] S. Benedetti, S. Buratti, A. Spinardi, S. Mannino and I. Mignani,
"Electronic nose as a non-destructive tool to characterise peach
accurate way. This required the improvement and cultivars and to monitor their ripening stage during shelf-
optimization of various parameters, like the thresholds or the life", Postharvest Biology and Technology, vol. 47, no. 2, 2008,
goal resolution. If these parameters were not optimal it would pp. 181-188.
result in extended running times or even cases where the [5] P. Ribeiro, S. Cardoso, A. Bernardino and L. Jamone, "Fruit
circularity could not be extracted, these cases will now be quality control by surface analysis using a bio-inspired soft tactile
discussed. sensor", 2020 IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), 2020, pp. 8875-8881.
Further modifications could be made to test and potentially [6] B. Anuja and A. Bansal, “Fruits and vegetables quality evaluation
optimize this algorithm, like changing the color extraction using computer vision: a review,” J. King Saud Univ. Comput.
Inf. Sci. 33, 2021, pp.243-257.
method, and by using Otsu’s method for image thresholding.

671

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.
 [7] Z. Yudong, S. Wang, G. Ji and P. Phillips, “Fruit classification [12] M. J. McDonnell, “Box-filtering techniques,” Computer
using computer vision and feedforward neural network,” Journal Graphics and Image Processing, vol. 17, no. 1, 1981, pp. 65–70.
of Food Engineering 143, 2014, pp.167-177. [13] T. -W. Chen, Y. -L. Chen, and S. -Y. Chien, “Fast image
[8] Z. Yudong and L. Wu, “Classification of Fruits Using Computer segmentation based on K-Means clustering with histograms in
Vision and a Multiclass Support Vector Machine,” Sensors HSV color space,” IEEE 10th Workshop on Multimedia Signal
(Basel, Switzerland) 12, 2012, pp.12489 - 12505. Processing, 2008, pp.322-325.
[9] Seema, A. Kumar and G. S. Gill, “Automatic Fruit Grading and [14] S. Biswas, and R. Hazra, “Robust Edge Detection Based on
Classification System Using Computer Vision: A Review,” Modified Moore- Neighbor,” Optik, 2018, 168, 931-943.
Second International Conference on Advances in Computing and [15] F. -J. Yang, "An Implementation of Naive Bayes Classifier,"
Communication Engineering, 2015, pp.598-603. 2018 International Conference on Computational Science and
[10] S. Naik and B. Patel, “Machine vision based fruit classification Computational Intelligence (CSCI), 2018, pp. 301-306.
and grading - a review,” International Journal of Computer [16] D.R. Sarvamangala, R. V. Kulkarni, “Convolutional neural
Applications 170, 2017, pp.22-34. networks in medical image understanding: a survey,” Evol. Intel.
[11] Y. Gong, B. Liu, X. Hou and G. Qiu, "Sub-window Box Filter," 15, 2022, pp.1–22.
2018 IEEE Visual Communications and Image Processing
(VCIP), 2018, pp. 1-4.

672

Authorized licensed use limited to: PES University Bengaluru. Downloaded on November 15,2024 at 14:15:56 UTC from IEEE Xplore. Restrictions apply.