DrBerry: Detection of Diseases in Blueberry Bush Leaves

Cristopher Morales, Edgar Cavero and Willy Ugarte

Department of Computer Science, Universidad Peruana de Ciencias Aplicadas, Lima, Peru

Keywords:

YoloV5, Computer Vision, Blueberry, Machine Learning, Plague, Disease.

Abstract:

The following research presents a mobile application that can recognize the following plages usually found on

blueberry leaves: oidium, heliothis and alternaria. These diseases affects the growth of the bush an thus reduce

its yield. Additionally, an open dataset will be available for future investigations. Yolov5, a convolutional

neural network, is used for the development of the model, data collection was performed in the Fundo San

Roberto, Huaral-Peru, and data augmentation techniques were used to increment the amount of workable data.

Thanks to this the following results were obtained: accuracy of 84% and recall of 91%. We predict that the

model could be improved to recognize other plages given the right amount of data.

1 INTRODUCTION

Our main goal is the detection of diseases in the leaves

of the blueberry bush, with the scientiﬁc name Vac-

cinium corymbosum.

This research topic was chosen because the blue-

berry is currently the main agricultural export prod-

uct of Peru, which gives it great economic importance

and motivates the care of crops from diseases that can

damage it.

For the development, the creation of a dataset of

images of blueberry leaves was proposed, since there

was no public use dataset for this plant at least on our

research. After the dataset was created, a YoloV5-

based model was trained in order to detect three dif-

ferent diseases (Oidium, Alternaria, Heliotis)

com-

monly present in blueberry leaves during the budding

stage.

Finally, with the trained model, a mobile applica-

tion will be created for Android devices which can

be used locally without the need to use the internet,

since many crops are not within the reach of mobile

networks.

For as long as man has been able to grow crops,

they have always been affected by the forces of na-

ture, be they weather problems or pests. Nowadays,

we have been able to learn to predict the weather quite

https://orcid.org/0000-0002-7510-618X

“Integrated Management of Pests and Diseases of

Cranberry Crops” (in spanish) - Bayer CropScience -

https://cdn.blueberriesconsulting.com/2016/12/manejo-

integrado-enfermedades-arandano.pdf

accurately thanks to various scientiﬁc devices, how-

ever, pests are still a problem to be tamed. Generally,

in Peru, technological advances are usually applied

years after they are developed and are often limited to

private property.

The last 5 years have seen major changes in the

ranking of products exported abroad within the agri-

cultural sector, with a new export leader that has man-

aged to impose itself on more traditional products

such as grapes, avocado and coffee.

This product is the blueberry and, at the time of

writing

, it has become the main agricultural product

exported from Peru.

This explosion in the export of the fruit is reﬂected

in a great source of income for Peru, so it is necessary

not to miss the opportunity to research technological

products that can contribute to reduce maintenance

costs and production losses.

The problem arises from the high demand for

blueberry production because, to the best of our

knowledge, there is very little work on the identiﬁ-

cation of blueberry leaf diseases, with the exception

of the private sector.

If a person with limited contacts or resources

wants to conduct a research, it will be difﬁcult to do

since they must obtain data, travel to blueberry or-

“The blueberry is positioned as the main national

agro-export product in the last twelve months” - Na-

tional Ministry of Commerce and Turism of Peru -

https://www.gob.pe/institucion/mincetur/noticias/672598-

el-arandano-se-posiciona-como-el-principal-producto-de-

agroexportacion-nacional-en-los-ultimos-doce-meses

Morales, C., Cavero, E. and Ugarte, W.

DrBerry: Detection of Diseases in Blueberry Bush Leaves.

DOI: 10.5220/0012207100003598

In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 1: KDIR, pages 355-364

ISBN: 978-989-758-671-2; ISSN: 2184-3228

355

chards and collect their own necessary information,

all assuming they are allowed access to the crops

since, to the best of our knowledge, there are not pub-

lic datasets available to use for training. It is for these

reasons that in a public way it is far from being solved.

Previous studies have been carried out for differ-

ent types of agricultural products such as (Yu et al.,

2023) soybeans, (Luaibi et al., 2021) citrus, (Gunisetti

et al., 2022) maize and (Amin et al., 2022) corn, in

which the use of both computer vision and convolu-

tional neural networks have been used to detect vari-

ous diseases.

However, these products have different pests from

which they are affected. In addition, although simi-

lar research has been done in Peru, according to our

search we did not ﬁnd a public use data-set for our

blueberry species to be treated in order.

The main components of our solution consist of

an image recognition application using YoloV5 that

is capable of being used on a mobile device without

the need for an Internet connection to make use of the

model.

In addition to the application, a dataset of blue-

berry bush leaf images will be created and made

freely available for future research. The limitations

of our work are temporal, since it was carried out in a

speciﬁc year, and technological, since the application

must run on mobile devices.

Our main contributions are as follows:

- We are developing a model in Yolov5, which al-

lows us to infer from an image diseases in the

leaves of the blueberry bush.

- We have constructed a dataset based on 3 diseases

of blueberry bush leaves.

- We have performed an experimental analysis with

the data collected.

This paper is distributed in the following sections:

First, we have summarized important papers about the

use of Yolov5 within the agricultural sector in Section

III, then we will explain some important concepts and

theorems about Yolov5 in Section IV. After that we

will explore in detail the experiments we carried out

during the development of the project in Section V

and ﬁnally, we show our main conclusions.

2 RELATED WORKS

It has become very popular in recent years the use of

computer vision and deep learning to ﬁnd solutions to

all kinds of problems, among these we ﬁnd the area

of disease detection in plants. We have found differ-

ent works that helped us understand the use of Yolo

while also inspiring us to try our own approach, from

these we can highlight the articles on modiﬁcations

on the base architecture of Yolo and the approach of

development in mobile devices.

In (Mathew and Mahesh, 2022) the authors pro-

pose a method for the detection of diseases in the pep-

per plant leaves that can be used in mobile phones.

For its implementation they decide to use Yolov5, fol-

lowing all the basics suggested when working with

Yolo, like getting the data-set, labeling in a manual

way and training the model. Before the training they

had to make different experiments to get the most ap-

propiate kind of model since Yolo have different mod-

els variants like nano, small and medium, these being

the ones suggested if the end goal is to deploy the

model on a cellphone. Like them, we have decided to

opt for the same model variants for our implementa-

tion since we also plan on use our model on mobile

devices. Thanks to them we have a better knowledge

about the use different versions models of Yolo and

will follow the suggested steps when working with it.

In (Liu et al., 2022) the authors propose a modiﬁed

architecture of the Yolov4 model, adding a triple at-

tention mechanism and the focal loss function, as well

as using the K-means method to group the data and

specifying a new anchor box. The ﬁrst change was

crucial since some pest in tomato are to small to no-

tice and by adding these modules they will have three

parallel branches to capture the data and get more fo-

cus on the feature channels. Contrary to them, since

the data we recollected was taken at short distance and

most of the diseases evaluated occupies enough pix-

els of the images, we opt to use the base Yolov5 model

without changing its layers.

In (Chen et al., 2022) the authors propose to add

new modules to the Yolov5 model, they replaced the

module C3 for an Involution-Bottleneck module. Af-

ter that, they add the SE module to the last layer of the

backbone network, and ﬁnally they replace the loss

function by Efﬁcient Intersection over Union in order

to minimize the discrepancy and accelerate the con-

vergence of the model. The ﬁrst change was useful

since they want to reduce the number of operations

through the learning process and get more focus on

long-distance information of the spatial range while

reducing the number of network parameters. By re-

placing the loss function they seeked to reduce the

discrepancies and accelerate the model convergence,

obtaining an average accuracy of 70 %, 5% higher

than traditional Yolo. We, on the other hand, are using

the base model of Yolov5 because most of our images

are in a short-distance, getting us a beneﬁt on model

convergence since we have less instances to infer on

images. As a result, our results are similar with an

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

356

Figure 1: Residual Boxes Generated in Image, from (Redmon et al., 2016).

average accuracy of 86% to our images.

In (Dai et al., 2022) the authors propose the de-

velopment of an hybrid model of Yolov5, they seek

for a model that can help them reduce time and com-

putational costs, for that reason they develop DA-

ActNN Yolov5, where Activation Compressed Train-

ing(ActNN) is used to compress the modules and pa-

rameters of Yolo, this is done without affecting the

model structure and the ﬁnal model performance an,

as a ﬁnal result, they got that the time consumption

time doesn’t exceed 30 minutes and the model runs

more efﬁciently in a 0.65% compared with Yolo. In-

stead, as one of our main contributions is the creation

of our open dataset, we decided to use the base Yolov5

structure in order to avoid missing our deadlines.

In (Delgado and Obeso, 2019) the authors made

a research in order to prove the effectiveness that a

machine learning solution would have on the iden-

tiﬁcation of plages and diseases that affects blue-

berry bushes in Trujillo, Peru. Their investigation

concluded that a computer vision-solution will im-

prove the chances of recognizing plagues in the berry

bush leaves. However, the research didn’t include

a database to work with, something that could have

helped in the development of such machine learning

solution. As part of our contributions we provide an

open dataset for future models and, since the investi-

gation was done on both the same type of crop and on

similar climate conditions, we can rest asure knowing

that our Yolov5 model will help in recognizing the

plagues of the berry bush leaves.

3 LEAF DISEASE IN

BLUEBERRIES BY INFERENCE

WITH Yolov5

3.1 Preliminary Concepts

In this section, we present the main concepts used in

the development of our work.

We seek to perform detection (inference) of blue-

berry bush leaf diseases, using computer vision and

convolutional neural networks, where these methods

help us to extract the necessary information from the

images of the leaves.

Deﬁnition 1 (Residual Boxes (Redmon et al., 2016)).

This is when an image is divided into grids of ﬁxed

dimensions, say n*n. Where each cells is trying to

detect object inside them.

Example 1 (Residual Boxes). We can see the residual

boxes generated in the Figure 1.

Deﬁnition 2 (Bounding Box Regression (Redmon

et al., 2016)). This marks an outline that highlights an

object in an image. Where their attributes are Height,

width, class and box center.

Example 2. We can see on the second step of the Fig-

ure 1 how some areas are highlight in yellow.

Deﬁnition 3 (Intersection Over Union (IoU) (Red-

mon et al., 2016)). This describes how the boxes are

overlap. Where if IoU is equal to 1 is meaning that

the predicted bounding box is the same that the real

box. This mechaninsm help us to eliminated bounding

boxes that are not equal to the real box.

Example 3. We can see on the third step of the Figure

1 how the image get more clear after deleting some

bounding boxes to get one with the most probability

score.

Deﬁnition 4 (YOLO Arquitecture (Redmon et al.,

2016)). It is a sequence of 3 components:

DrBerry: Detection of Diseases in Blueberry Bush Leaves

357

Figure 2: Object Detector Arquitecture, from (Bochkovskiy

et al., 2020).

- Backbone: This refers to a type of neural network

called a convolutional neural network, which is

designed to gather and combine visual character-

istics from images at various levels of detail.

- Neck: This refers to the use of a sequence of layers

in a neural network to blend and merge visual fea-

tures extracted from images, in order to propagate

them through the network and make a prediction

based on them.

- Head: This refers that uses the features from

neck and uses them to make predictions, includ-

ing the location of object and their corresponding

classes.

Example 4. We can see on the Figure 2, how the ar-

chitecture is represented and follow different steps un-

til get the prediction.

3.2 Method

In this section, the main contributions proposed will

be explain and detailed.

3.2.1 Blueberry Bush Leaf Diseases Dataset

Our ﬁrst contribution is the creation of a dataset that

was used to train the Yolov5 model, which contains 4

classes: three of diseases and one of healthy leaves.

There is a total of 439 images containing 1 single

disease and 105 images with 2 or more diseases.

The picture were taken at the Fundo San Roberto,

located at the Sayan district, Huaura Province, in the

Lima region.

The climate was sunny the picture were taken dur-

ing midday, the cameras used were a Huawei Y7 2019

cellphone, a Samsung A30 and a Sony CyberShot

DSC-W630.

• Image Pre-Processing: After the collection of the

images, they received a manual ﬁlter where those

that contained very blurry or distorted images

were discarded in order to guarantee the quality

of the images. Later, different data augmentation

techniques were used in the images in order to in-

crease the number of total images; For this, we

used the image data generator, from the Keras li-

brary in Python, in order to reduce a possible over-

ﬁtting problem and increase the diversity of our

Figure 3: Examples of Diseases Alternaria (a) ,Heliothis

(b), Oidium (c).

data, as shown in Figure 3. We have also made a

reshape to the images to reduce the number of pix-

els in order to reduce computational costs when

training the model. Then, we apply a combination

of data augmentation techniques such as: rotation

transform, to rotate the image to horizontal an ver-

tical ﬂip with a parameter of 0.6; brightness trans-

formation, where 0.0 means minimum brightness

and 1.0 means maximum brightness; channel shift

transformation, where a channel shift range of 60

is applied; ﬁnally the ﬁll mode reﬂect is used to

ﬁll in the proximity pixels.

• Labeling Images Using Roboﬂow: Continuing the

pre-processing, we have used the Roboﬂow tool

and its online manual labeling tool as shown in

Figure 4, which gives us a compressed ﬁle con-

taining our data and labels in folders organized

and ready to use in Yolo.

3.2.2 Architecture

Our second contribution for this investigation is the

use of computer vision and convolutional neural net-

works in our plague detection model using the Yolov5

framework.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

358

Figure 4: Roboﬂow Label Tools berry leaves an example.

In Figure 5 we can observe the workﬂow, it all

starts by taking a picture of a blueberry bush leaf.

- Take a photo: We take a picture of the leaf that we

want to analyze with our program, it will be saved

on our image gallery.

- Upload Image: From the app we search for the

picture that we took and select it, its format can

be in JPEG or PNG.

- Scaling Image: Once the image is loaded it will

be scaled to 1024x1024 px.

- Residual Boxes: Our detection model divides the

image into an array of square cells in a N*N for-

mat, where each cell in the array is given the task

of locate and predict the classes inside the area

it covers with its respective probability and conﬁ-

dence values.

- Bounding Box Regression: This step consists of

determining the bounding boxes that corresponds

to the rectangles that highlight all the objects in

the image, it is possible to have as many bounding

boxes as there are objects in the given image. Yolo

will determine the bounding boxes using a unique

regression model which follows the following for-

mat, where Y represents the ﬁnal vector of each

bounding box.

Y = [pc, bx, by, bh, bw, c1, c2] (1)

– pc: represent the probability value of the grid

containing an object.

– bx,by : represents the x and y coordinate values

of the center of the bounding box.

– bh,bw : represents the height and width values

of the bounding box

– c1,c2 : represents the classes found in the im-

ages, we can have as many classes as the case

requires. In our case there are 4.

- Intersection Over Union: Within our images, mul-

tiple candidate grid boxes may appear for the pre-

diction, although not all of them are relevant. That

is the objective of the IOU, where the less relevant

grid boxes are discarded to keep the relevant ones.

4 EXPERIMENTS

In this section we will talk about the experiments car-

ried out, as well as everything necessary to replicate

them, and the results obtained from the experiments

will also be discussed.

4.1 Experimental Protocol

In this subsection we will detail the development en-

vironment conﬁgurations in which the experiments

were carried out, as well as the hardware and software

used.

Finally, the step-by-step will be detailed to carry

out your own training.

A computer has been used as the development

platform, the operating system was Ubuntu 22.04.2

and the Pytorch framework and Yolov5 environment

were created in Anaconda.

The program was coded in Python 3.10.6 and the

cuda version was 12.1. For the hardware, an AMD

Ryzen 7 5700x CPU processor, NVIDIA Geforce

RTX 3070 GPU and 32 GB of RAM at 3200 Mhz

have been used.

The speciﬁc conﬁgurations are found in the Table

4.1.1 Training a Custom Yolov5 Model

• Setting up the Yolo enviroment: We must clone

the Yolov5 repository from Github (Jocher, 2020).

Yolov5 runs on Torch, where it can easily be run

on Google Colab or Kaggle. Cloning creates a

ﬁle named Yolov5 on the machine, where it will

contain Yolo’s speciﬁc directory structure.

• Setting up data and directory structure: In or-

der to label and manage the structure of the ﬁles

quickly and easily, the Roboﬂow tool has been

used where, after uploading and labeling our im-

ages, we can download a zip ﬁle which will con-

tain the ﬁles necessary for our training.

• Training the model: Training the model is done by

running the train.py ﬁle from the notebook. Pa-

rameters such as number of epochs, batch size,

image size, among others, can be established.

Once executed, a folder will be generated which

will contain the weights of the trained model. To

carry out the detection of diseases we must use

the detect.py ﬁle. After training, the best.pt ﬁle

DrBerry: Detection of Diseases in Blueberry Bush Leaves

359

Figure 5: Framework Flow.

Table 1: Development environment setting.

Parameter Conﬁguration

Operating system Ubuntu 22.04.2

Deep learning framework PyTorch 1.21.1

Programming language Python 3.10.6

GPU accelerated environment CUDA 12.1

GPU Geforce RTX 3070 8gb

CPU AMD Ryzen 7 5700x @3.4 Ghz

Figure 6: Healthy leaf detected.

located in runs/train/exper/weights/best.pt is ob-

tained. The Figure 6 is an example of the images

that are generated after the training has ﬁnished.

4.2 Results

In this subsection, we will talk about the experiments

carried out and the results obtained in each one of

them will be detailed.

4.2.1 Model Evaluation Metrics

To achieve the tests or evaluations of the model for the

detection of Alternaria, Heliothis, Oidium and healthy

leaves, an exhaustive analysis has been carried out

on the performance of the model, and other methods

have also been discussed, such as the following indi-

cators: Precision (P), Recall (R), F1-Score (F1), Ac-

curacy (A) and Average Precision (AP).

4.2.2 Analysis of Data Augmentation Results

Initially, we want to check if applying data augmen-

tation techniques to our blueberry leaf dataset is ben-

eﬁcial for training our Yolov5s model.

This test evaluates the model using the mAP, pre-

cision and loss value metrics, in Figure 7a and Fig-

ure 7b, which show the train values, validation loss

as well as the precision and mAP50 values at each

epoch and, when comparing both versions, a better

recognition percentage is observed in the DrBerry

model when using data augmentation techniques to

the dataset.

The validation and training loss values decrease

over time, as well as showing some points of conver-

gence, as well as giving us an emphasis on a possi-

ble overﬁtting since the train/loss values decrease and

val/loss increase as shown in Figure 7a and Figure 7b.

The curves represented in Figure 7a and Figure 7b

show us a great difference when using data augmenta-

tion methods, obtaining better results in both accuracy

and mAP from the start of the training.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

360

(a) Original Data.

(b) Augmented Data.

Figure 7: Train Loss (a), Validation Loss (b), Accuracy rate (c), mAP50 rate (d).

It can also be seen that the accuracy value ranges

remained high until the end of the training.

As shown in Table 2, tests were performed with

three groups of data augmentation techniques applied

to the DrBerry dataset and trained with the Yolov5s

model.

The results were compiled to evaluate the perfor-

mance of each combination in order to ﬁnd the most

favorable one.

All the trainings were carried out with a batch

size of 8, with 300 epochs, in some cases the

intervention of the early stopping mechanism was

recorded, and with all the default learning rates that

can be found in the route ./yolov5/data/ hyps/hyp.no-

augmentation.yaml.

Number 2 presents the use of 3 data augmentation

techniques with which a value of 81.8% accuracy was

obtained while Number 3 presents the use of 4 data

augmentation techniques where we obtain an 86.4%

accuracy, a 4.6% improvement compared to the pre-

vious one.

On the other hand, Number 4, with 5 data augmen-

tation techniques, presents a value of 79.5% accuracy

and, although the latter shows a worse performance,

it shows us that modifying this value is not recom-

mended for the type of images we are working with.

The augmentation of data has helped us to obtain

signiﬁcant improvements when using Yolov5s model

compared to when we did not use it.

4.2.3 Analysis of Data Augmentation and

Modiﬁed Hyperparameters

In order to validate the performance of the hypepa-

rameters in conjunction with the data augmentation

techniques applied, the following Figure 8 is shown

where the legends of data original, data v1,data v2

and data v3 refer to to the techniques used in Table

The results of the experiments show us consider-

able improvements between the use of improved hy-

perparameters against the results that do not use them,

where the results obtained in Hyp-Low Augmented

managed to reach values of 88%, 94% and 94% in

accuracy, recall and mAP50 respectively.

These are signiﬁcant values since we were able

to increase 5%, 24% and 14% compared to the data

without increased hyperparameters.

Likewise, we highlight the MOSAIC hyperparam-

eter, which helps us to have more variability in our

images Figure 9, which helps us to reduce overﬁtting

as shown in Figure 10.

This can be seen in tables (C) and (D), where in

the ﬁrst case these values are not converging but rather

they are increasing over time, contrary to the second

case where the values decrease over time until they

converge.

DrBerry: Detection of Diseases in Blueberry Bush Leaves

361

Table 2: Results obtained from training Yolov5s with different data augmentation methods.

Rotation

Range

Horizontal

Flip

Vertical

Flip

Brightness

Channel

Shift

Diseases Precision% Recall% mAP50%

Alternaria

Heliothis

Oidium

Healthy

All

62.4

66.8

83.7

63.2

69.0

81.6

36.5

34.3

25.6

44.5

77.8

58.0

44.9

33.0

53.4

X X X

Alternaria

Heliothis

Oidium

Healthy

All

67.5

84.4

88.4

86.7

81.8

63.3

57.4

76.5

86.1

70.8

54.4

78.7

90.2

94.0

79.3

X X X X

Alternaria

Heliothis

Oidium

Healthy

All

84.8

88.4

89.4

82.9

86.4

68.5

61.0

82.1

82.7

73.6

76.7

79.3

92.8

89.8

84.7

X X X X X

Alternaria

Heliothis

Oidium

Healthy

All

82.1

89.0

81.6

65.3

79.5

88.8

71.7

76.5

65.4

75.6

88.9

80.1

81.8

64.8

78.9

Figure 8: Training for Yolov5s with various kinds of data

augmentation techniques and types of hyperparameters.

4.2.4 Analysis of Yolov5s and Yolov5m Using

Data Augmentation and Modiﬁed

Hyperparameters

Table 3 shows the comparison of the results obtained

after carrying out different experiments, in this case

Number 3 of Table 2 has been used as data for our

evaluations, where we obtain similar results between

precision, recall and mAP50, highlighting the in-

crease of 5%, 0.1% and 2% respectively in the M ver-

sion of the Yolov5 model.

Figure 9: Result applying the MOSAIC hyperparameter.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

362

Figure 10: Yolov5 S No Hyp vs Hyp Low.

Table 3: Results obtained from training with Yolov5s and Yolov5m utilizando Hyp Low in Testing Images.

Model Weigth Diseases Precision % Recall % mAP50 %

Yolov5s 14.11Mb

Alternaria

Heliothis

Oidium

Healthy

All

81.4

86.9

97.9

87.5

88.4

90.4

88.4

98.7

92.8

92.6

90.9

92.2

99.4

94.1

Yolov5m 40.61Mb

Alternaria

Heliothis

Oidium

Healthy

All

87.4

96.4

98.8

89.3

93.0

87.9

92.7

98.7

91.6

92.7

90.6

98.0

99.2

96.7

96.1

5 CONCLUSIONS AND

PERSPECTIVES

In this work we managed to create a Yolov5 model

capable of identifying three different types of plagues

that afﬂicts the blueberry bush’s leaves, reaching a

94.1% and 96.1% mAP50 as shown in the Table 3.

We also point out that our model presents a high

Recall percentage, value that we consider important

in the identiﬁcation of diseases in the plant’s ear-

lier stages, since a misidentiﬁcation may result in

the plant being considered healthy, when that is not

the case, left unchecked and potentially damaging its

growth.

We also want to point out the importance of the

dataset and the data-augmentation techniques we ap-

plied, being able to collect data is not an easy task

when the source is not easily available and it can re-

DrBerry: Detection of Diseases in Blueberry Bush Leaves

363

sult in a suboptimal dataset, thus, by applying data

augmentation we managed to improve our results as

it has been previously discussed, and other works try

different approaches for similar problems (Rodr

ıguez

et al., 2021; Leon-Urbano and Ugarte, 2020; Lozano-

Mej

ıa et al., 2020).

For future works we have thought about taking

two possible routes: i) Making an recomentadion

software that, using our current model, gives the user

useful information about the disease detected, what

potential harm it could have on the blueberry bush if

left unchecked and posible solutions.

It would be an ofﬂine solution since some planta-

tions lacks the internet connectivity. ii) Improve our

model making it able to detect more diseases, by the

time of this writing new versions of Yolo have been

developed and it would be interesting to see what ben-

eﬁts they can bring to our proyect.

We could also explore different data augmentation

techniques to further improve our dataset for future

trainings.

REFERENCES

Amin, H., Darwish, A., Hassanien, A. E., and Soliman, M.

(2022). End-to-end deep learning model for corn leaf

disease classiﬁcation. IEEE Access, 10:31103–31115.

Bochkovskiy, A., Wang, C., and Liao, H. M. (2020).

Yolov4: Optimal speed and accuracy of object detec-

tion. CoRR, abs/2004.10934.

Chen, Z., Wu, R., Lin, Y., Li, C., Chen, S., Yuan, Z., Chen,

S., and Zou, X. (2022). Plant disease recognition

model based on improved yolov5. Agronomy, 12.

Dai, G., Hu, L., and Fan, J. (2022). Da-actnn-yolov5: Hy-

brid yolo v5 model with data augmentation and acti-

vation of compression mechanism for potato disease

identiﬁcation. Computational Intelligence and Neuro-

science, 2022.

Delgado, R. and Obeso, G. (2019). Soluci

on de machine

learning en el reconocimiento de plagas para plan-

tones de ar

andano.

Gunisetti, L., Koduri, S. B., and Jagannathan, V. (2022).

Optimized deep learning system for smart maize leaf

disease detection in iot platform via routing algorithm.

Multimedia Tools and Applications.

Jocher, G. (2020). Yolov5 by ultralytics.

Leon-Urbano, C. and Ugarte, W. (2020). End-to-end elec-

troencephalogram (EEG) motor imagery classiﬁcation

with long short-term. In IEEE SSCI, pages 2814–

2820.

Liu, J., Wang, X., Miao, W., and Liu, G. (2022). Tomato

pest recognition algorithm based on improved yolov4.

Frontiers in Plant Science, 13.

Lozano-Mej

ıa, D. J., Vega-Uribe, E. P., and Ugarte, W.

(2020). Content-based image classiﬁcation for sheet

music books recognition. In 2020 IEEE EIRCON.

Luaibi, A. R., Salman, T. M., and Miry, A. H. (2021). De-

tection of citrus leaf diseases using a deep learning

technique. International Journal of Electrical and

Computer Engineering, 11.

Mathew, M. P. and Mahesh, T. Y. (2022). Leaf-based dis-

ease detection in bell pepper plant using yolo v5. Sig-

nal, Image and Video Processing, 16:841–847.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time ob-

ject detection. In IEEE CVPR.

Rodr

ıguez, M., Pastor, F., and Ugarte, W. (2021). Clas-

siﬁcation of fruit ripeness grades using a convolu-

tional neural network and data augmentation. In IEEE

FRUCT, pages 374–380.

Yu, M., Ma, X., and Guan, H. (2023). Recognition method

of soybean leaf diseases using residual neural network

based on transfer learning. Ecological Informatics,

76:102096.

KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval

364