=Paper=
{{Paper
|id=Vol-3687/Paper_3.pdf
|storemode=property
|title=Software for UAV Images Processing for Object Identification
|pdfUrl=https://ceur-ws.org/Vol-3687/Paper_3.pdf
|volume=Vol-3687
|authors=Kateryna Merkulova,Yelyzaveta Zhabska,Ivan Ivanenko
|dblpUrl=https://dblp.org/rec/conf/dsmsi/MerkulovaZI23
}}
==Software for UAV Images Processing for Object Identification==
https://ceur-ws.org/Vol-3687/Paper_3.pdf
Software for UAV Images Processing for Object Identification
Kateryna Merkulova, Yelyzaveta Zhabska and Ivan Ivanenko
Taras Shevchenko National University of Kyiv, Volodymyrska str. 64/13, Kyiv, 01601, Ukraine
Abstract
This paper describes the research and comparative analysis of methods for object identification
in UAV images with the aim of determining the most relevant method in accordance with the
described quality criteria in the context of detecting different types of vehicles for its further
use during software implementation.
Three identification methods were chosen for the study, namely methods based on ResNet,
MobileNet and EfficientDet models. During the research, three quality criteria for evaluating
identification methods were developed and described.
As a result, none of the methods showed the best results for all three quality criteria, therefore
priorities were set for each quality criterion. Having evaluated the results of the quality criteria
for each of the researched identification methods, while taking into account the priorities of
the quality criteria, it was concluded that the method based on the MobileNet model is the most
optimal among the researched methods in the context of vehicle identification on UAV images.
Keywords 1
Object identification, UAV, vehicles, artificial neural network, ResNet, MobileNet,
EfficientDet, comparative analysis
1. Introduction
With the growing use of unmanned aerial vehicles (UAVs) in fields ranging from military
surveillance to geodesy and environmental monitoring, image processing research is becoming a
necessary component for the effective use of these technologies [1, 2]. One of the most significant areas
of research is the identification of vehicles through the analysis of images obtained from UAVs. This
article explores the methods and technologies used to process UAV imagery to accurately identify
different types of vehicles.
Knowledge and understanding of modern image processing methods for the purpose of vehicle
identification is becoming increasingly important for maintaining safety, efficiency and sustainable
development of society. The results of these studies depend not only on the development of the latest
technologies, but also on providing our world with greater safety and efficiency in various areas of life.
Solving the problem of automating the process of detecting suspicious objects on UAV images
during martial law in Ukraine is an important aspect of security and control in the state. That is why the
developed software is designed to identify different types of vehicles on UAV images. The discussed
topic is quite relevant at the moment, as drones have become an important tool in the process of waging
war. And that is why operative identification of suspicious vehicles on UAV images is a very urgent
task in today’s realities for our country.
2. Related works and research objective
Nowadays, there are many software applications for UAV images processing with the purpose of
object identification. Applications of this type are used in various fields of human activity, starting with
agriculture and ending with the military industry. Today, the most popular software tools for UAV
images processing with the purpose of object identification usually include the following programs:
Dynamical System Modeling and Stability Investigation (DSMSI-2023), December 19-21, 2023, Kyiv, Ukraine
EMAIL: kate.don11@gmail.com (K. Merkulova); y.zhabska@gmail.com (Y. Zhabska); super-ivan-ivanenko@knu.ua (I. Ivanenko)
ORCID: 0000-0001-6347-5191 (K. Merkulova); 0000-0002-9917-3723 (Y. Zhabska); 0009-0008-0658-9801 (I. Ivanenko)
©️ 2023 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
CEUR
Workshop
ceur-ws.org
ISSN 1613-0073
25
Proceedings
1. Pix4Dmapper – the main functionality of the program includes automatic processing of images
taken from different angles using photogrammetry and computer vision algorithms. The program
supports images in JPEG, TIFF, PNG and RAW formats, which allows it to work with most cameras
and UAVs on the market [3]. The program is able to determine various parameters of the territory, such
as area, volume, height, slope, length, etc. Using the deep learning technology, the program can
automatically recognize and classify objects in images, such as buildings, roads, trees and others.
2. DroneDeploy is a cloud-based software for processing aerial images taken by UAVs to create
detailed maps and 3D models. It is a full-featured platform for flight mission planning, data collection
and image processing, which allows to quickly and easily create map data and 3D models of various
objects [4]. DroneDeploy’s core functionality includes flight mission planning using an interactive map
and automatic UAV control.
3. AgroScout is a software for processing images from a UAV, designed for diagnosing plant
diseases and assessing the condition of crops in real time [5]. This technology is used in agriculture to
monitor and diagnose plant diseases in the early stages, which allows to avoid the spread of diseases
and preserve the harvest. AgroScout uses computer vision technology to analyze images obtained from
drones.
Having analyzed the existing software solutions, it can be concluded that the analogs in question are
complex software products that are used in their subject area, namely:
1. Pix4Dmapper – the functionality is focused on the analysis of the earth surface.
2. DroneDeploy – functionality is based on the creation of detailed maps and 3D models.
3. AgroScout – the functionality consists in diagnosing plant diseases and assessing the condition
of crops.
It is obvious that developing a software product with similar functionality is not the best solution,
since there will be no question of novelty. That is why the developed software will be intended for a
slightly different subject area, namely for the identification of different types of vehicles in images
taken from UAVs. The discussed topic is quite relevant at the moment during martial law in Ukraine,
as drones have become an important tool for conducting various war operations. Therefore, it is
important to implement the software that can perform identification of suspicious vehicles on UAV
images as quickly and accurate as possible.
3. Methods of research
Nowadays, there is no exact analytical solution to the problem of object identification in images,
which complicates the development of a universal algorithm. Nevertheless, in order to endow computer
systems with the possibility of so called “vision”, a large number of methods and algorithms have been
created and proposed [6, 7]. The purpose of this section is to consider the most popular of them, after
which, on the basis of tests of their software implementations, select one or another method for the
implementation of the final software product. Based on the previous experience, as well as on
information from various articles, the following most popular and used methods for identifying objects
in the image were selected:
4. MobileNet is a set of architectures of deep neural networks [8]. They are optimized to run on
devices with limited computing resources.
5. ResNet is an abbreviation for “Residual Network” (Network with reverse connections). It was
developed in order to ensure successful training of deep neural networks, avoid the problem of
gradient damping and facilitate learning [9].
6. EfficientDet is a family of object detector architectures that combine efficiency and high
accuracy [10]. They were developed in order to provide efficient image processing and object
recognition with minimal computational costs.
So, at the moment, identification methods have been determined, which were selected for further
implementation. Each of the selected methods is a certain type of artificial neural network, therefore,
for their implementation, it is necessary to conduct their training on a certain set of data. Today, the
issue of national security during martial law is extremely important, which is why it is extremely
important to timely identify suspicious vehicles in the front-line territories and beyond. Thus, these
types of objects for identification in UAV images are quite relevant now. Therefore, it was decided to
use a variety of vehicles as objects for identification in the images taken from the UAV. The next task
26
is the formation of data for training artificial neural networks. For now, there are more than a dozen sets
of images with annotations to them, which would be suitable for the task of this research, have already
been created. As a result of detailed analysis, a dataset called VisDrone, which contains 8408 annotated
images, was selected for training [11].
VisDrone is a dataset designed for the tasks of object detection and tracking in images captured by
unmanned aerial vehicles (drones). This dataset contains annotations for various object classes such as
pedestrians, cars, trucks, buses, motorcycles, and others. Within the VisDrone dataset, each object is
annotated with a bounding box that shows its position in the image. These annotations are used to train
and evaluate object detection models. Figure 1 shows an image from the VisDrone dataset along with
the bounding boxes contained in the image annotation.
Figure 1: Image from VisDrone and its annotation
TensorFlow [12] is used as a tool for implementing selected models. TensorFlow is an open source
machine learning and deep learning software developed by Google. It provides a framework for
building and training a variety of artificial intelligence models, such as neural networks.
In the process of training, monitoring was also carried out, which reflects the effectiveness of model
training. Monitoring displays such loss functions as Classification loss and Localization loss.
Classification loss is a loss function used during the training of a neural network for classification.
It measures the distance between predicted and actual class labels and helps train the network during
the training process [13, 14]. Localization loss is a loss function used for training of a neural network
to localize objects in images. It measures the distance between the predicted and actual object
coordinates and helps train the network during the training process.
Figures 2-4 present the graphs of the loss functions described above for each of the trained models.
Figure 2: ResNet model training process
To compare the training results of the selected models, Table 1 is presented, which contains the
values of the loss functions for each model at the end of training.
Before analyzing the obtained results, it is necessary to clarify that the lower the value of the loss
function, the better the trained model performs the task. That is, the smaller the loss function, the more
the model predictions correspond to the expected result.
27
Figure 3: MobileNet model training process
Figure 4: EfficientDet model training process
Table 1
Results of model training
classification_loss localization_loss
ResNet 0.2723 0.1988
MobileNet 0.2344 0.1638
EfficientDet 0.2259 0.0064
Thus, analyzing the results of model training, it can be concluded that the EfficientDet network
performed best in the training process, as it has the lowest values of the loss functions. It should be
noted that the value of classification_loss in all three models is almost at the same level, while
localization_loss has a much smaller value in the EfficientDet model. So, the selected models were
implemented and the results of their training were compared. Now it is necessary to check the
performance of the implemented methods in practice. It can be done with the help of the proposed
quality criteria, which will be discussed later.
At this stage, the question arises, what should be paid attention to when choosing quality criteria?
In the research process, they will be used for vehicle identification methods in UAV images, therefore,
based on the previous experience and information from various papers, the following criteria will be
considered the most appropriate for this type of method.
3.1.1. The ratio of the number of correctly identified objects to the number of
all objects
The ratio of the number of correctly identified objects to the number of all objects is a criterion that
evaluates such a quantitative characteristic of the model as the ability to identify given objects. In order
not to specify the name of this metric every time, let's highlight for it, for example, the symbol R. The
expression 1 demonstrates the formula for calculating the quality criterion R for the identification
method, using a sample of images of size N:
∑𝑁𝑖=1 𝑚𝑖
𝑅= 𝑁 , (1)
∑𝑖=1 𝑘𝑖
where R is the quality criterion of the identification method, namely the ratio of the number of correctly
identified objects to the number of all objects, N is the number of all images, mi is the number of objects
28
that the method was able to correctly identify on the i-th image, ki is the actual number of objects in the
i-th image.
In which case is it considered that the method correctly identified the object in the image? The
quality assessment metric called Intersection over Union, which will be discussed in more detail later
in the paper, will help answer the first question. The value of the IoU quality assessment metric can
vary from 0 to 1. It is generally accepted that IoU > 0.5 is a good prediction of the object detector,
otherwise the prediction is not good. Thus, when counting correctly identified objects, it is needed to
calculate the value of the IoU quality assessment metric for each of them. If IoU > 0.5, the object is
included in the calculation, otherwise it is ignored. Expression 2 demonstrates the formula for
calculating the number of correctly identified objects in the i-th image:
𝑘𝑖
1, 𝐼𝑜𝑈𝑗 > 0.5,
𝑚𝑖 = ∑ { (2)
0, 𝐼𝑜𝑈𝑗 ≤ 0.5,
𝑗=1
where mi is the number of correctly localized objects in the i-th image, ki is the actual number of objects
in the i-th image, IoUj is the value of the IoU quality assessment metric calculated for the j-th object in
the i-th image.
Thus, when calculating the quality criterion R, formula 2 will be substituted into formula 1.
The last question, that remains open within the scope of this point, is what should be the number of
all images in order to calculate the value of the quality criterion with a given error ε. Expression 3
demonstrates the formula for finding the number N for a given error ε:
𝜀 = |𝑓(𝑛 + 𝑠𝑡𝑒𝑝) − 𝑓 (𝑛)|, (3)
where ε is the specified error for calculating f, f(N) is the value of the metric f for some object identifier,
calculated using a sample of images of size N, n is the current value of the number of images for
calculating the value of f, step is a fixed step that increases the value n for each subsequent iteration.
Figure 5 contains a block diagram of the algorithm for finding the value of N for a given error ε.
There are probably no separate rules or advice for choosing the permissible error value ε, again
everything depends on the intuition of the developer himself. Of course, it’s not worth to take too large
values for ε, such as 10-2. Based on the experience of previous research of various computational
methods, it is recommended to take the value of ε as 10-5. Usually this accuracy is completely sufficient
for most numerical methods.
3.1.2. Intersection over Union
Intersection over Union is a criterion that evaluates the accuracy with which the model localizes
objects in the image. Expression number 4 demonstrates the formula for calculating the IoU metric:
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑂𝑣𝑒𝑟𝑙𝑎𝑝
𝐼𝑜𝑈 = , (4)
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑈𝑛𝑖𝑜𝑛
where IoU is the Intersection over Union quality assessment metric, Area of Overlap is the area of
intersection of the predicted bounding frames with the actual bounding frames, Area of Union is the
area of the union of the predicted bounding frames with the actual bounding frames.
From formula 4 it can be seen that the possible values for the IoU metric are in the numerical range
from 0 to 1, taking into account the extreme points. An IoU > 0.5 is generally considered to be a good
predictor of an object detector. The next formula will allow to calculate the average value of the IoU
quality criterion for a sample of N images:
∑𝑁𝑖=1 𝐼𝑜𝑈𝑖
𝐼𝑜𝑈𝑐 = , (5)
𝑁
where IoUc is the average value of the Intersection over Union quality assessment metric, IoUi is the
value of the IoU quality assessment metric calculated for the i-th image and N is the number of all
images.
29
Figure 5: Block diagram of the algorithm for finding N-value for a given ε-value
3.1.3. Average object localization time
Average object localization time is a criterion that is designed to demonstrate the speed of the
researched method. The following expression demonstrates the calculating of the average value of the
T criterion.
∑𝑁𝑖=1 𝑇𝑖
𝑇= , (6)
𝑁
where T is the average object identification time in the image, Ti is the average object identification
time in the i-th image and N is the number of all images.
Thus, all three criteria cover the most significant characteristics of the object identification method,
namely identification ability, localization accuracy, and recognition speed. It should also be noted that
each of the mentioned metrics will be calculated not for a single image, but for some sample of size N,
that is an averaged value is used.
30
4. Research results
The results of the calculation of all three quality criteria for each object identification method are
summarized in Table 2.
Table 2
Quality criterions for object identification methods
R IoUс T
ResNet 0,82798 0,8767 0.00282
MobileNet 0,92156 0,8656 0.00245
EfficientDet 0,62243 0,8349 0.00087
Now, based on the obtained results, comparative analysis of the selected identification methods can
be conducted. For this, their qualitative and quantitative comparison was carried out in the context of
the previously described quality criteria.
Let’s start with such a quality criterion as Intersection over Union. The physical meaning of this
metric is a numerical representation of the accuracy with which the identification method predicts the
location of the object in the image. Figure 6 demonstrates three graphs for each of the identification
methods, which show the dependence of the average value of the IoU quality criterion on the number
of images required for its calculation.
Figure 6: Graphs of the IoU(N) function for three identification methods
Qualitative comparison of detection methods according to the IoU metric: starting from about 50
images, the values of the IoU metric for all three identification methods actually do not change, that is,
they acquire constant values. For the ResNet model the final value is IoU1 ≈ 0.87673, for the MobileNet
model IoU2 ≈ 0.86559 and for EfficientDet IoU3 ≈ 0.83493. In general, the IoU metric can take values
between 0 and 1, including extreme values. A detection method for which IoU > 0.5 is considered good.
Therefore, based on this fact, it can be stated that all three methods for which the calculations were
performed are quite good detectors. But, of course, one of them showed a slightly better result than the
others, it is a model based on ResNet.
Quantitative comparison of identification methods by the IoU metric can be calculated as follows:
0,87673 − 0,86559
𝑃𝐼𝑜𝑈 = · 100% ≈ 1,3%, (7)
0,86559
31
0,87673 − 0,83493
𝑃𝐼𝑜𝑈 = · 100% ≈ 5%. (8)
0,83493
As a result of the quantitative comparison of the two detection methods based on the IoU metric, it
can be concluded that the ResNet model predicts the placement of given vehicles on the image by 1.3%
more accurately than the MobileNet model, and by 5% more accurately than the EfficientDet model.
Next, for the comparative analysis of the methods, we will take the metric R, that is, the ratio of the
number of correctly located vehicles to the number of all vehicles. The physical meaning of this metric
is a numerical representation of the ability of the method to correctly identify the given vehicles. Figure
7 shows contains graphs for each of the identification methods, which show the dependence of the
numerical value of the metric R on the number of images required for its calculation.
Figure 7: Graphs of the R(N) function for three identification methods
Qualitative comparison of detection methods according to the R metric: starting from about 125
images, the values of the R metric for the three identification methods do not actually change, that is,
they acquire constant values. In Figure 7 it can be seen that in the case of the method based on
MobileNet the final value R1 ≈ 0.92156, for the method based on ResNet R2 ≈ 0.82798, for EfficientDet
R3 ≈ 0.62243. Summarizing the qualitative comparison of the methods, it can be definitely concluded
that the identification method based on the MobileNet model is the best among the studied methods
according to the quality criterion R. At the next stage, it is necessary to determine how much the
MobileNet model is better than the other studied methods in the context of the identification of given
vehicles by the R metric.
Quantitative comparison of identification methods by the R metric can be demonstrated with the
following:
0,92156 − 0,82798
𝑃𝑅 = · 100% ≈ 11,3%, (9)
0,82798
0,92156 − 0,62243
𝑃𝑅 = · 100% ≈ 48%. (10)
0,62243
As a result of the quantitative comparison of the detection methods by the R metric, it can be
concluded that the method based on the MobileNet model has 11.3% more ability to identify vehicles
in the image than the method based on ResNet, and 48% more than the method based on EfficientDet.
The last quality criterion for comparing the identification methods is the average identification time
of one vehicle in the image. Figure 8 shows three graphs for each of the identification methods, which
show the dependence of the numerical value of the metric T on the number of images N required for its
calculation.
Qualitative comparison of detection methods by the metric T: starting from about 125 images, the
values of the metric T for all three identification methods practically do not change, that is, they acquire
32
constant values. For the method based on the ResNet model, the final value is T1 ≈ 0.00282 s, for the
method based on MobileNet T2 ≈ 0.00245 s, for EfficientDet T3 ≈ 0.00087 s. In the case of the first
two metrics, which were considered earlier, the method in which the corresponding metric has a larger
numerical value was considered better. Instead, for this metric, the method in which the numerical value
is smaller will be better. This is intuitive because the shorter the average identification time of one
vehicle in the image, the more efficient the identification method is.
Figure 8: Graphs of the function T(N) for three identification methods
After a qualitative comparison of the methods, it can be unequivocally stated that the method based
on EfficientDet is the best among the investigated methods according to the quality criterion T. Now it
is necessary to determine how much this method is better than the others in numerical equivalent.
For their quantitative comparison, it will be easier to find how many times one of them is more than
the other. This can be calculated as follows:
0,00282 0,00245
≈ 3,24; ≈ 2,91. (11)
0,00087 0,00087
Therefore, as a result of the quantitative comparison of the three identification methods based on the
T metric, it can be concluded that the identification method based on EfficientDet on average identifies
one vehicle in the image 3 times faster than the other two methods.
5. Conclusion
After conducting a comparative analysis of the studied identification methods, it is currently difficult
to determine which of them is the best in the context of vehicle identification in UAV images, since
each of them showed the high result in at least one quality criterion.
Thus, based on the obtained results, it can be said that each of the considered identification methods
has its own advantages and disadvantages. However, it is necessary to clearly determine which of the
given methods is the most optimal according to the given criteria in the context of vehicle identification
on UAV images. Since none of the methods showed the best results for all three quality criteria, it is
now necessary to prioritize each quality criterion.
Based on the expert experience of the authors, it can be noted that the quality criterion T, which
characterizes the speed of the method, is the least significant (third priority) in this context, since each
of the methods showed results of the order of 10-3 seconds. That is, it will be difficult to notice a
significant difference between them in terms of speed in practice. Instead, other quality criteria that
characterize other aspects of the studied methods are more significant, as they can be felt in practice.
Quality criterion R, which characterizes the ability of the method to identify vehicles, is the most
significant (first priority). As it will be clearly seen that one method managed to identify 10 vehicles
and the other only 7 in the same image. Accordingly, only the second priority remains for the IoU
quality criterion. Having evaluated the results of the quality criteria for each of the researched
33
identification methods, while taking into account the priorities of the quality criteria, it was concluded
that the method based on the MobileNet model is the most optimal among the researched methods in
the context of vehicle identification on UAV images. Because it showed the best results for the quality
criterion R, which has the highest priority, while breaking away from the other two methods by a margin
(by 11.3% - ResNet; by 48% - EfficientDet). The next most important is the IoU quality criterion,
according to which the method based on the MobileNet model showed the second result, lagging behind
the ResNet by 1.3%. In other words, MobileNet and ResNet are actually equal to each other according
to this criterion. According to the third least significant quality criterion, the MobileNet-based method
showed again the second result, which is three times worse than the first result, which was shown by
the EfficientDet-based method. Again, it may seem that this is quite a noticeable difference, but if we
take into account the fact that the obtained speed values are of the order of 10 -3 seconds, then in practice
this difference will not be noticeable. Although EfficientDet is the fastest in the context of vehicle
identification in UAV images, on the other hand it performed the worst in two other more significant
quality criteria. Therefore, based on the obtained results, the method based on the EfficientDet model
is the worst in the context of vehicle identification in UAV images according to the given quality
criteria, taking into account their priority.
6. References
[1] A. N and U. S V, “Drone Object Detection Using Deep Learning Algorithms,” 2021 Third
International Conference on Inventive Research in Computing Applications (ICIRCA),
Coimbatore, India, 2021, pp. 1187-1192, doi: 10.1109/ICIRCA51532.2021.9544983.
[2] G. Cheng, Z. Shao, Q. Cheng and H. Luo, “Small Object Detection on Drone-view Images from
Multiple UAVs Combination,” 2023 30th International Conference on Geoinformatics, London,
United Kingdom, 2023, pp. 1-7, doi: 10.1109/Geoinformatics60313.2023.10247874.
[3] Pix4Dmapper. URL: https://www.pix4d.com/product/pix4dmapper-photogrammetry-software/
[4] DroneDeploy. URL: https://www.dronedeploy.com/
[5] AgroScout. URL: https://agro-scout.com/
[6] V. Martsenyuk, O. Bychkov, K. Merkulova and Y. Zhabska, “Exploring Image Unified Space for
Improving Information Technology for Person Identification,” in IEEE Access, vol. 11, pp. 76347-
76358, 2023, doi: 10.1109/ACCESS.2023.3297488.
[7] S. Kumar and C. Kumar, “Drone/UAVs Surveillance feeds Video Object Recognition Using
Transfer Learning,” 2022 International Conference on Futuristic Technologies (INCOFT),
Belgaum, India, 2022, pp. 1-4, doi: 10.1109/INCOFT55651.2022.10094328.
[8] A. G. Howard, et al., “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
Applications”, CoRR, 2017, abs/1704.04861. URL: http://arxiv.org/abs/1704.04861.
[9] K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual Learning for Image Recognition,” 2016 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016,
pp. 770-778, doi: 10.1109/CVPR.2016.90.
[10] M. Tan and A. Yu, “EfficientDet: Towards Scalable and Efficient Object Detection”, April 15,
2020. URL: https://ai.googleblog.com/2020/04/efficientdet-towards-scalable-and.html
[11] P. Zhu, et al., “Detection and Tracking Meet Drones Challenge” in IEEE Transactions on Pattern
Analysis & Machine Intelligence, vol. 44, no. 11, pp. 7380-7399, 2022. doi:
10.1109/TPAMI.2021.3119563.
[12] TensorFlow. URL: https://www.tensorflow.org/
[13] Ivanytska, L. Zubyk, Y. Zubyk and D. Ivanov, “The advertising prediction model based on
machine learning technologies”, CEUR Workshop Proceedings of the 8th International Conference
"Information Technology and Interactions" (IT&I-2021), Kyiv, Ukraine, December 1-3, 2021.
CEUR-WS, vol. 3179, pp. 35-44. https://ceur-ws.org/Vol-3179/Paper_4.pdf
[14] Kutyrev, A., Kiktev, N., Kalivoshko, O., Rakhmedov, R. Recognition and Classification Apple
Fruits Based on a Convolutional Neural Network Model. CEUR Workshop Proceedings of the 9th
International Conference "Information Technology and Implementation" (IT&I-2022), Kyiv,
Ukraine, November 30 - December 2, 2022. CEUR-WS, vol. 3347, pp. 90-101. https://ceur-
ws.org/Vol-3347/Paper_8.pdf
34