LIDL4Oliv: A Lightweight Incremental Deep Learning Model for

Classifying Olive Diseases in Images

Emna Guermazi

1,2,3

, Afef Mdhaffar

1,3

, Mohamed Jmaiel

1,3

and Bernd Freisleben

ReDCAD Laboratory, ENIS, University of Sfax, B. P. 1173 Sfax, Tunisia

National School of Electronics and Telecommunications of Sfax, University of Sfax, 3018 Sfax, Tunisia

Digital Research Center of Sfax, 3021 Sfax, Tunisia

Department of Mathematics and Computer Science, Philipps-Universit

at, Marburg, Germany

Keywords:

Olive Disease Detection, Knowledge Distillation, Incremental Learning.

Abstract:

We present LIDL4Oliv, a novel lightweight incremental deep learning model for classifying olive diseases

in images. LIDL4Oliv is ﬁrst trained on a novel annotated dataset of images with complex background.

Then, it learns from a large-scale deep learning model, following a knowledge distillation approach. Finally,

LIDL4Oliv is successfully deployed as a cross-platform application on resource-limited mobile devices, such

as smartphones. The deployed deep learning can detect olive leaves in images and classify their states as

healthy or unhealthy, i.e., affected by one of the two diseases “Aculus Olearius” and “Peacock Spot”. Our

mobile application supports the collection of real data during operation, i.e., the training dataset is continu-

ously augmented by newly collected images of olive leaves. Furthermore, our deep learning model is retrained

in a continuous manner, whenever a new set of data is collected. LIDL4Oliv follows an incremental update

process. It does not ignore the knowledge of the previously deployed model, but it (1) incorporates the current

weights of the deployed model and (2) employs ﬁne-tuning and knowledge distillation to create an enhanced

incremental lightweight deep learning model. Our conducted experiments show the impact of using our com-

plex background dataset to improve the classiﬁcation results. They demonstrate the effect of using knowledge

distillation in enhancing the performance of the deployed model on resource-limited devices.

1 INTRODUCTION

Agricultural crops have a signiﬁcant nutritional role

in our lives. However, plant diseases threaten agricul-

tural yields and pose a considerable risk to the produc-

tion of food. Indeed, about 20% of global crop losses

stem from plant diseases (Savary et al., 2012; Ney

et al., 2013). This percentage has increased in the past

decade due to the inﬂuence of pollution and shifts in

climate patterns. To tackle this problem, researchers

have made signiﬁcant efforts to identify and detect

plant disease at an early stage. The proliferation of

smart farming technologies has fostered the develop-

ment of innovative technologies to automatically de-

tect plant diseases, such as smartphone-assisted plant

disease diagnosis.

To build an effective mobile device solution for

plant disease detection, several researchers have pro-

posed to use appropriate deep learning models, ei-

ther (a) by hosting a deep learning model on a cloud

server and implementing a mobile application that in-

teracts with the deep learning model using a com-

munication protocol, or (2) by embedding the deep

learning model directly into a mobile application run-

ning on a mobile device. The ﬁrst approach supports

the use of large-scale deep learning models running

on a server with substantial storage capacity and ex-

tensive computational resources. Moreover, this ap-

proach is not suitable for real-time agricultural appli-

cations. The second approach supports real-time agri-

cultural data analysis, but creating and deploying a

lightweight high-quality neural network is not an easy

task. Several researchers have created automated mo-

bile solutions that leverage the power of deep learn-

ing across various plant types. Some of these solu-

tions are specialized to detecting diseases in tomato

leaves (Aishwarya et al., 2023; Chandan et al., 2022).

Others identify diseases in corn plants (Hidayat et al.,

2019). Additionally, there are efforts to detect ter-

restrial plants in the Philippines using similar tech-

niques (Valdoria et al., 2019). Furthermore, some

researchers proposed solutions for disease detection

Guermazi, E., Mdhaffar, A., Jmaiel, M. and Freisleben, B.

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images.

DOI: 10.5220/0012466900003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 2, pages 583-594

ISBN: 978-989-758-680-4; ISSN: 2184-433X

583

of a variety of plants (Ahmed and Reddy, 2021; Gu-

rav et al., 2022). However, all these approaches do

not take into account the need to adapt the deployed

model to the changes seen in real-life images. More-

over, up to this point, a smart deep learning based mo-

bile application speciﬁcally designed to detect olive

leaf diseases is still absent. Detecting olive leaf dis-

eases is particular relevant for the Mediterranean and

Aegean regions that collectively contribute to about

95% of the world’s olive production (Soyyigit and

Yavuzaslan, 2018). This justiﬁes the need for a smart

deep learning-based mobile application dedicated to

detecting olive diseases.

In this paper, we propose a novel lightweight in-

cremental mobile deep learning model to detect and

classify olive diseases in images. For classiﬁcation,

we use the object detection model YOLOv5-n to de-

tect an olive leaf in an image and then proceed with

disease classiﬁcation. We also create and annotate a

complex background dataset. This dataset improves

the capabilities of the deep learning model to rec-

ognize diseases under various settings and in images

with different realistic backgrounds. We use knowl-

edge distillation (KD) for object detection to enhance

the performance of YOLOv5-n. Next, we deploy our

lightweight neural network in a mobile application

dedicated to olive farm management and disease pre-

diction. Our approach supports sustainable data col-

lection and offers an incremental update process. This

process enables the model to always be up-to-date to

the features observed in real-life images.

This paper is structured as follows. Section 2 dis-

cusses related work. In Section 3, we present our ap-

proach, called LIDL4Oliv. Section 4 sketches imple-

mentation details. Section 5 presents the conducted

experiments and discusses the obtained results. Sec-

tion 5 concludes the paper and outlines some areas for

future research.

2 RELATED WORK

Detecting plant diseases is among the primary appli-

cations of smart agriculture. Deep learning models

are widely used for the detection and classiﬁcation of

plant leaf diseases.

Several studies have shown a dedicated interest

in the development of deep learning models for the

recognition of olive leaf diseases. (Alshammari et al.,

2023) presented an approach to analyze olive leaves.

The authors used the Whale Optimization Algorithm

(WOA) to choose necessary features. Then, they clas-

siﬁed the leaves using an optimized artiﬁcial neural

network (ANN). (Alshammari and Alkhiri, 2023) pre-

sented an approach that combines a recurrent neu-

ral network architecture with an ant colony optimiza-

tion algorithm. The ant colony optimization algo-

rithm was used to extract relevant features. Then, dis-

ease classiﬁcation was performed using RNN models.

(Osco-Mamani and Chaparro-Cruz, 2023) employed

a modiﬁed VGG16 architecture for the classiﬁcation

of olive leaf diseases in Tacna. (Mamdouh and Khat-

tab, 2022) used a deep learning model based on the In-

ceptionV3 architecture. (Ksibi et al., 2022) presented

MobiRes-Net, a neural network that combines the ar-

chitectures of ResNet50 and MobileNet. Although

these methods propose efﬁcient deep learning mod-

els to classify olive leaf diseases, the complexity of

the developed deep learning models hinders their de-

ployment on resource-limited mobile devices, such as

smartphones. Moreover, none of these methods in-

volves a mobile application-based approach to detect

olive diseases automatically.

Research on mobile application approaches that

offer automated deep learning-based detection of

plant diseases can be classiﬁed into two categories.

The ﬁrst category uses cloud servers to host deep

learning models (Ahmed and Reddy, 2021; Chan-

dan et al., 2022; Hidayat et al., 2019), while the

second category involves deploying a deep learning

model within the mobile application itself (Aishwarya

et al., 2023; Gurav et al., 2022; Valdoria et al., 2019).

(Chandan et al., 2022) developed a plant disease de-

tection system. This system integrated an Internet-

of-Things (IoT) platform to collect various measure-

ments, including humidity and temperature. Then,

these measurements were uploaded to a cloud server

for processing. Leaf images were also transmitted to

the cloud platform for analysis. Then, the authors em-

ployed a mobile app to display measurement values,

send SMS alerts, and provide notiﬁcations. (Hidayat

et al., 2019) developed an Android application for di-

agnosing corn diseases using deep learning. Here,

farmers select an image from their gallery, which is

then sent to a cloud server for processing. A convo-

lutional neural network (CNN) predicts the disease,

and the diagnosis is subsequently displayed in the mo-

bile application. (Ahmed and Reddy, 2021) proposed

a cross-platform mobile application designed to de-

tect infected plants. The mobile app enables taking

a picture of the leaf via the camera or uploading an

existing image from the smartphone. Then, a cloud

server uses a CNN model for disease detection. Sub-

sequently, the diagnosis of the disease is displayed

in the app. The methods belonging to the ﬁrst cate-

gory successfully achieve automatic plant disease de-

tection, but their approaches necessitate the availabil-

ity of extensive computing resources and cannot be

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

584

Complex-background

dataset

Initial lightweight

model

Deployment on

mobile-app

Data-collection

Incremental update

process

Large-scale

model

Triggers

Incremented model

Figure 1: LIDL4Oliv.

easily deployed on resource-limited mobile devices.

The second category of approaches addresses this

issue by developing a lightweight neural network and

deploying it directly in a mobile application. (Valdo-

ria et al., 2019) proposed a mobile application named

“iDahon”. This app aims to improve terrestrial dis-

ease classiﬁcation and diagnosis in the Philippines us-

ing a deep learning model. (Aishwarya et al., 2023)

suggested using a custom CNN model to classify

tomato leaf diseases. The application uses a Flask

API to establish communication between the user in-

terface and the underlying model. (Gurav et al., 2022)

presented an Android application that offers different

functionalities. First, it enables farmers to buy and

sell saplings. Moreover, it enables the plants’ dis-

ease detection via employing a lightweight Tensor-

Flow Lite based deep learning model. It also en-

ables the farmers to establish an agricultural com-

munity. Additionally, it offers up-to-date weather

forecasts for the duration of one week and delivers

real-time GPS tracking by utilizing satellite informa-

tion sourced from OpenWeatherMap. Although the

methods in the second category effectively integrate

lightweight neural network models into mobile ap-

plications, they fail to take into account the models’

need to accommodate changes occurring in real-life

images.

In contrast to the previously mentioned state-of-

the-art methods, we present an incremental update

process that allows us to maintain an updated deep

learning model that is well-adapted to the features of

real-life images. Furthermore, we ensure sustainable

data collection, as opposed to other mobile disease di-

agnosis solutions.

3 LIDL4Oliv

Our approach, called LIDL4Oliv, is based on an incre-

mental lightweight deep learning model that detects

and classiﬁes olive leaf diseases via a cross-platform

mobile app. Figure 1 illustrates LIDL4Oliv (i.e.,

Lightweight Incremental Deep Learning model for

classifying Olive diseases). Our approach involves

ﬁve phases. The ﬁrst phase includes several steps:

(i) Initially, we identify our deep learning model that

will be used for deployment. (ii) Then, we create

our complex background dataset. (iii) Next, we iden-

tify a large-scale deep learning model that is used as

our teacher model. (iv) Subsequently, we use knowl-

edge distillation to optimize the performance of our

student model. In the second phase, we deploy our

lightweight deep learning model as a cross-platform

mobile application. The third phase consists of con-

tinuously enriching our original dataset by newly pro-

cessed images. This ensures the creation of a sustain-

able dataset of olive leaf images. Upon accumulat-

ing 30% of new data relative to the original dataset,

the fourth phase is initiated where the deep learning

model undergoes an incremental update process to

enhance its performance, guaranteeing that it is con-

stantly adapted to the new features of the collected

images. In the ﬁfth phase, we deploy our incremented

deep learning model in the mobile app. Each phase is

further detailed in the following sections.

3.1 Deep Learning Model Identiﬁcation

Classiﬁcation of olive diseases in images is a chal-

lenging task. In this paper, we use olive leaves to de-

tect and identify olive diseases. To solve this task,

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images

585

(a) Aculus Olearius (b) Aculus Olearius (c) Aculus Olearius (d) Peacock Spot

(e) Peacock Spot (f) Peacock Spot (g) Healthy (h) Healthy

Figure 2: Different olive leaf states (Uguz,S., 2020).

we employ an object detection model instead of a

classiﬁcation model. This decision is motivated by

three reasons. First, real-life images usually contain

different backgrounds. Thus, it is more suitable to

make a distinction between the background and the

region of interest itself (i.e., the olive leaf) before

identifying the disease. Object detection ﬁrst local-

izes the object and then proceeds with classiﬁcation.

By following this process, we have a higher chance

of classifying the disease. Second, deploying a single

model and using it for both localization and classiﬁ-

cation is more practical and resource-conserving than

deploying multiple models (initially a deep learning

model to eliminate the background, followed by an-

other model to classify the leaf disease). Third, an

object detection model can detect multiple objects per

image. However, in this case, we use it on images

where each image includes only one leaf. This is

due to the non-existence of an open real-life anno-

tated olive leaf dataset containing multiple leaves per

image. Thus, we decided to primarily create a dataset

where each image includes a background and solely

one leaf. This serves as the ﬁrst set of training data

to train our initial model. Then, after deploying the

model as a mobile app, it is possible to collect realis-

tic images and annotate them on the basis of experts’

knowledge. Subsequently, we are able to adjust our

model accordingly to localize and classify multiple

leaves in an image.

We conducted a search for annotated olive leaf

images and identiﬁed an open dataset established by

(Uguz,S., 2020). This dataset was created for olive

leaf disease classiﬁcation. It involves olive leaf im-

ages that correspond to three states (i.e., ‘Healthy” or

affected with “Aculus Olearius” or “Peacock Spot”).

However, it is composed of images having a simple

and unrealistic background, as illustrated in Figure 2.

Thus, relying on such a dataset would yield a deep

learning model that is not suitable for real-life olive

leaf disease detection. Given these circumstances, we

used the Uguz dataset (Uguz,S., 2020) and generated

a complex background dataset with bounding box an-

notations to localize a leaf and identify its state. The

next section explains the process of generating the

complex background dataset with bounding box an-

notations.

We use the YOLO architecture as our object de-

tection model. Indeed, contrary to two-stage detec-

tors like Faster R-CNN, YOLO’s architecture enables

us to detect multiple objects in a single pass through

the neural network, making it faster. Additionally,

YOLO achieves a good balance between speed and

accuracy, qualifying it for usage scenarios requiring

real-time processing. Moreover, YOLO is available

in different lightweight architectures, such as YOLO-

s and YOLO-n, which make it adequate for our use

case. We identiﬁed YOLOv5-n, following the work

of (Ye et al., 2023), as our model to use for deploy-

ment.

3.2 Generation of an Annotated

Complex Background Dataset

To create a dataset including leaf images with a com-

plex background, we followed two steps. First, we

constructed a dataset with a white background and

created the corresponding bounding box labels. Then,

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

586

we used an overlapping technique to generate im-

ages with complex background. These intricate back-

grounds correspond to images captured in the garden,

incorporating elements such as sand, hands, etc. Al-

though these backgrounds may not perfectly match

the context of photographing in the ﬁeld, they allow

the model to detect the leaf within images containing

backgrounds. Thus, our model differentiates between

the background and the leaf, and identiﬁes the state of

the leaf accurately. This dataset allows us to deploy

an initial model in real life (i.e., ready to be used by

farmers).

3.2.1 Generation of an Annotated White

Background Dataset

Our newly produced white background dataset is

based on Uguz’s dataset (Uguz,S., 2020). It incorpo-

rates bounding box annotations to train deep learning

models for object detection and classiﬁcation, con-

trary to Uguz’s dataset that is solely used for im-

age classiﬁcation. The original dataset is composed

of two folders: one for training and the second one

for testing. Within each folder, there are three sub-

folders. Each of these subfolders contains images that

correspond to a speciﬁc class of an olive leaf. This

class can be either “Healthy” or affected with either

”Aculus Olearius” or ”Peacock Spot”. In the remain-

der of this paper, we refer to the “bounding box an-

notations” by “annotations”, and to the true labels of

each class by “ground truth labels”.

To construct our dataset, we ﬁrst applied data aug-

mentation techniques to Uguz’s dataset. This allowed

us to introduce greater image variations.

Then, we used the Canny method (Canny, 1986)

as a contour detection algorithm to annotate 31%

(1,479 images) of our dataset. We also used several

preprocessing techniques such as Gaussian blur, his-

togram equalization, and median ﬁltering to create an-

notations. We developed a Tkinter (Steen Lumholt

and Guido van Rossum, 2009) interface to support the

selection of images with accurate bounding boxes, as

shown in Figure 3.

The Tkinter interface is used for validating the an-

notation of each image. For the rest of the images that

were not annotated, we used YOLOv5-m to complete

the annotation process. We employed the annotated

images, which represent 31% of the dataset, to train

the model. Then, we used the trained model to pro-

duce the annotations of the remaining images. YOLO

allowed us to annotate around 66% of the dataset, i.e.,

leaving 3% (i.e., 150 images) not annotated. The re-

maining images were manually annotated, via the use

of LabelImg (Tzutalin, 2015), as shown in Figure 4.

We selected LabelImg due to its user-friendly simplic-

Figure 3: The Tkinter interface used to validate the bound-

ing box annotations.

ity.

Figure 4: Number of images annotated using different

methods.

3.2.2 Overlapping Process

The second step involves changing the white back-

ground of the dataset to a more complex one. To

guarantee the retention of the established annotations,

we undertook a meticulous overlapping process. This

process involves two steps. The ﬁrst one consists of

generating masks from the white background dataset

using a pre-trained U2-Net model (Qin et al., 2020).

The generated masks are binary images where the

background is represented by black pixels, and the re-

gion of interest, which is the leaf, is represented by

white pixels. The employed U2-net model was pre-

trained on the DUTS dataset (Wang et al., 2017). Af-

ter the masks’ generation, we moved on to the sec-

ond step. This step involved combining the gener-

ated masks with other images that represent a com-

plex background. The resulting complex background

dataset enables the YOLO model to accurately detect

and classify the state of the olive leaves in intricate

and varied settings, paving the way for improved clas-

siﬁcation performance in real-life situations.

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images

587

(a) Splash screen (b) Prediction page (c) Prediction results (d) Trees’ monitoring page

Figure 5: Selection of some pages of the mobile application.

3.3 Initial Training Process

To enhance the performance of our lightweight neu-

ral network, i.e., YOLOv5-n, we incorporated knowl-

edge distillation into the training process. This tech-

nique is based on knowledge transfer from a large-

scale neural network referred to as teacher to a

lightweight neural network named student to boost

its performance. Its initial application was in image

classiﬁcation tasks (Hinton et al., 2015; Gou et al.,

2023; Tian et al., 2020), subsequently expanding to

encompass object detection in images (Chen et al.,

2017; Kruthiventi et al., 2017). The integration of

knowledge distillation into object detection tasks sig-

niﬁcantly boosted the performance of the employed

models (Li et al., 2023; Yadikar et al., 2023). How-

ever, it is still a challenging task.

Some approaches are based on global distillation

where the student tries to mimic the whole feature

maps extracted from the middle layers of the teacher.

However, relying on all the feature maps might lead

to learning irrelevant knowledge that would not help

the student in its task. Therefore, some researchers

showed interest in local feature distillation (Wang

et al., 2019) to arrive at detector distillation, propos-

ing the distillation of backbone features extracted

from local regions surrounding objects. We use this

approach to distill knowledge from YOLOv5-m to

YOLOv5-n.

3.4 Cross-Platform Mobile Application

We deployed our model as a cross-platform mobile

application (i.e., compatible with both Android and

iOS). This application is designed to enable automatic

olive leaf disease prediction while also establishing an

olive farm management system. It is called “Olive

Leaf Disease Detector”, as illustrated in Figure 5a.

The splash screen of the application allows users to ei-

ther explore or sign in. It provides them with multiple

functionalities based on their roles. A user can have

the role of an admin, a farmer, or an expert. An admin

assumes responsibility for managing other users and

the existing trees on the property. Indeed, the admin

adds trees to the system. Each added tree corresponds

to an existing tree in the ﬁeld. Our application en-

ables the farmer to predict the state of an olive leaf

based on the prediction of a deployed deep learning

model. A farmer can either take a picture using the

camera of his or her smartphone or import it from his

or her gallery. Then, (s)he chooses the tree to which

the leaf is associated. The prediction page is shown in

Figure 5b. Subsequently, (s)he can either accept the

model’s prediction or seek an expert’s evaluation, as

illustrated in Figure 5c. Additionally, the farmer can

recommend an expert to the admin through a provided

form within the mobile application. The expert is re-

sponsible for answering the farmers’ requests and giv-

ing his or her opinion regarding the condition of spe-

ciﬁc leaves. Furthermore, (s)he is able to go through

all the leaves existing in the database via the mobile

application, enabling him or her to offer insights on

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

588

Student1

Dataset1

KD: Fine-grained

feature imitation

Dataset2

Fine-Tune

Student2

Fine-Tune

KD: Fine-grained

feature imitation

Teacher3

Fine-Tune

Student3

Fine-Tune

KD: Fine-grained feature

imitation

Incremental Update Process

Dataset3

Teacher1

Teacher2

Figure 6: Incremental update process.

any leaf status. When a leaf has two predictions, one

provided by the deep learning model and the other

given by an expert, the status attributed to it will be

based on the expert’s opinion.

Furthermore, our application enables the monitor-

ing of the health of the existing trees. By default, all

trees are considered to be in a healthy state. How-

ever, when a leaf state is changed to “Aculus Olear-

ius” or “Peacock Spot”, the status of that correspond-

ing tree changes to sick. After the identiﬁcation of an

unhealthy tree, it receives proper treatment. Then, a

farmer can access the mobile application and change

the status of the tree to healthy again. (S)he can also

delete all its leaves since they hold no meaning after

its status has changed. Furthermore, the farmer can

also delete the leaves after the season changes. Figure

5d presents the trees’ monitoring page. This appli-

cation also ensures a sustainable data collection. In-

deed, whenever a leaf status is predicted, the image is

saved with its corresponding label and bounding box

annotations in a remote server. Figure 7 illustrates

the communication process between the mobile app

screens (i.e., front-end) and the back-end, while also

representing the data collection process.

3.5 Incremental Update Process

To ensure the adaptability of our deployed model to

the features of the real-life images, we propose to use

an incremental update process. This process gives the

model the ability to evolve after gathering new data.

It involves a number of iterations as illustrated in Fig-

DL-powered

mobile app

Atlas Mongo DB

Distant Server

Node js Backend

Server

Data Collection

Figure 7: Global overview of the mobile application.

ure 6. Each iteration is initiated when the images of

new data reach 30% of the initial data. We retrain the

insights of the pre-trained teacher and the pre-trained

student by incorporating their pertained weights into

the update process. In each iteration, we ﬁne-tune the

pre-trained teacher on dataset

i+1

, which represents

the newly gathered dataset, at iteration

. Thus, we ob-

tain a ﬁne-tuned teacher. This teacher is referred to

as teacher

i+1

. We also apply ﬁne-tuning on the pre-

trained student. The produced student is referred to as

student

i+1

. Then, we apply the same knowledge dis-

tillation method that we used in the initial step (Wang

et al., 2019) (i.e., ﬁne-grained feature imitation). This

incremental update process reaches its end when a

student

i+1

(i.e., produced in iteration

) achieves good

performance when evaluated on a dataset

i+2

. It means

that this student model is well adapted to the features

of real-life images. This incremental process will be

implemented and experimentally validated upon col-

lecting real-life images through the mobile app and

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images

589

Table 1: Performance of YOLOv5-m: Setting 1 vs Setting 2.

Precision Recall mAP50-95

Setting1 Setting2 Setting1 Setting2 Setting1 Setting2

All 0.978 0.986 0.984 0.991 0.961 0.979

Healthy 0.988 0.987 0.989 0.994 0.955 0.972

Aculus Olearius 1 0.985 0.975 0.987 0.968 0.982

Peacock Spot 0.946 0.987 0.987 0.993 0.961 0.982

annotating them accordingly.

4 IMPLEMENTATION ISSUES

We implemented LIDL4Oliv using the SGD opti-

mizer with a momentum value of 0.9. The initial

learning rate, weight decay, and batch size were set

to 0.01, 0.0005, and 16, respectively. The training

and testing phases were conducted on an NVIDIA

RTX3080 GPU. For a fair comparison, all models un-

derwent training for 100 epochs.

5 EXPERIMENTAL EVALUATION

5.1 Experimental Settings

To evaluate our approach, we conducted a series of

preliminary experiments allowing us to identify our

experimental setting. These preliminary experiments

consist of training and evaluating the performance

of our teacher (i.e., YOLOv5-m) on the white back-

ground dataset, in two settings. Setting 1 involves

training the model from scratch, while in Setting 2 we

use pre-trained-model’s weights as a starting point.

These weights correspond to the YOLOv5-m model

after being trained on the COCO dataset (Lin et al.,

2014).

The main objective of this experiment is to iden-

tify the best setting to be used for the rest of the ex-

periments.

Table 1 and Figure 8 show the obtained results

when training YOLOv5-m in Setting 1 and Setting

2. Table 1 demonstrates that the overall precision, re-

call, and average mean precision (i.e., mAP50-90) are

higher when using Setting 2. These results also re-

veal that the recall and the mAP50-95 values of each

class individually exhibit higher values in compari-

son to Setting 1. However, the precision values of

the “Healthy” and the “Aculus Olearius” classes are

higher when using Setting 1. Figure 8 presents the

accuracy of predicting each class correctly. It shows

that Setting 2 shows a better classiﬁcation accuracy

for the “Healthy” and the “Aculus Olearius” classes,

Figure 8: Classiﬁcation accuracy of YOLOv5-m: Setting 1

vs Setting 2.

while, the accuracy of predicting the class “Peacock

spot” is the same for both settings. Thus, we can con-

clude that Setting 2 enables the model to achieve bet-

ter results in most of the evaluation metrics.

As a result, we work with Setting 2 in the rest of

this paper.

5.2 Evaluation of LIDL4Oliv

To evaluate the performance of LIDL4Oliv, we fol-

lowed three steps. First, we assessed the performance

of the pre-trained teacher model (i.e., trained on the

white background dataset) on images with complex

backgrounds. Second, we trained and evaluated the

performance of a teacher model trained on our com-

plex background dataset. Third, we evaluated the

performance of the student model with and without

knowledge distillation.

5.2.1 White Background Dataset

The results of the experiments conducted in the previ-

ous section show that YOLOv5-m, our teacher model,

achieves good results when evaluated on white back-

ground images. However, when we evaluate its per-

formance on images that have complex backgrounds,

we notice that its performance decreases, as illus-

trated in Table 2. YOLOv5-m achieves an overall

average mean precision of 0.298, demonstrating how

inefﬁcient the pre-trained model can be when faced

with images containing complex backgrounds. Fur-

thermore, we identify the limitation of using a white

background dataset for training and also justify the

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

590

(a) Confusion matrix of YOLOv5-n (w/o) KD (b) Confusion matrix of YOLOv5-n with KD

Figure 9: Confusion matrix of YOLOv5-n with and (w/o) KD.

Table 2: Results of YOLOv5-m trained on a white back-

ground dataset and tested on a complex background images.

mAP50 mAP50-95

All 0.394 0.298

Healthy 0.347 0.266

Aculus Olearius 0.609 0.465

Peacock Spot 0.225 0.163

need to use a complex background.

5.2.2 Complex Background Dataset

Training on a dataset with a complex background is

of high importance. This is because a real-life im-

age usually includes a complex background. Thus,

to ensure the performance of our deep learning net-

work in the ﬁeld, it is essential to train it on a similar

dataset. We use the same images when testing both

the teacher and the student. Samples of these images

are presented in Figure 11. Table 3 presents the ob-

tained results when training YOLOv5-m on our com-

plex background dataset.

Table 3: Results of YOLOv5-m trained and tested on com-

plex background dataset.

Class Precision Recall mAP50 mAP50-95

All 0.992 0.988 0.993 0.969

Healthy 0.989 0.989 0.994 0.959

Aculus Olearius 1 0.983 0.995 0.974

Peacock Spot 0.987 0.993 0.991 0.975

It is plausible that the achieved outcomes exhibit

remarkable performance in terms of precision, recall,

and mAP50-90. Indeed, its precision and recall val-

ues are 0.992 and 0.988, respectively. Also, its mean

average precision, mAP50-90, is equal to 0.969. Fig-

Figure 10: Confusion matrix of YOLOv5-m.

ure 10 presents the confusion matrix of YOLOv5-

m. It demonstrates that our teacher model achieves

reasonable classiﬁcation results across all designated

classes.

5.2.3 Comparative Performance Analysis of

YOLOv5-n With and Without Knowledge

Distillation

To evaluate the performance of our student model

(YOLOv5-n) in detecting and classifying olive dis-

eases, we compare its performance with and without

using knowledge distillation.

We also present the performance of the teacher

model (YOLOv5-m) for a meaningful comparison

with the performance of the student model. Figure 9

presents the confusion matrix of our student model

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images

591

Figure 11: Detection and classiﬁcation results of YOLOv5-n trained with KD.

Table 4: Performance of YOLOv5-n: With KD vs. without

KD.

YOLOv5-m

YOLOv5-n

(w/o KD)

YOLOv5-n

(with KD)

Precision 0.992 0.987 0.989

Recall 0.988 0.976 0.98

mAP50 0.993 0.994 0.994

mAP50-95 0.969 0.954 0.951

with and without KD. It is clear that knowledge dis-

tillation has improved the performance of the student

in classifying leaf diseases. Using KD enhanced the

student’s ability in classifying the ”Healthy“and the

”Peacock Spot“ state by an additional 1% for each.

The experimental results shown in Table 4 present

the precision, the recall, and the mean average pre-

cision of the teacher (YOLOv5-m), and the student

(YOLOv5-n trained with and without KD). It demon-

strates that YOLOv5-n trained using KD outperforms

YOLOv5-n trained without KD, in terms of precision

and recall. Speciﬁcally, YOLOv5-n trained using KD

obtains a precision value of 0.989 and a recall value of

0.980, while YOLOv5-n trained without KD achieves

a precision value of 0.987 and a recall value of 0.976.

We can also observe that the precision and the recall

of YOLOv5-n trained using KD achieves comparable

results when comparing it with the precision and the

recall values of the teacher YOLOv5-m, showing the

impact of ﬁne-grained feature imitation knowledge

distillation approach in enhancing the performance

of the student’s performance. However, we notice

that the mean average precision, mAP50-90, achieved

by YOLOv5-m trained without KD is slightly higher

than that achieved by YOLOv5-m trained with KD.

This is quite plausible and acceptable. Examples

of detection and classiﬁcation results of YOLOv5-n

trained with KD are shown in Figure 11. The test im-

ages are part of our generated dataset.

6 CONCLUSION

In this paper, we presented a novel lightweight incre-

mental deep learning model, called LIDL4Oliv, to de-

tect and classify olive leaf diseases in images. It is

based on YOLOv5-n for high-quality classiﬁcation.

We used knowledge distillation to achieve very good

results and deployed our model as a cross-platform

mobile application. This mobile application supports

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

592

automatic olive leaf detection and serves as a farm

management system. Our mobile application enables

a farmer to ask for an expert opinion if (s)he wants

further evaluation of the leaf status. We established a

sustainable data collection process by gathering pro-

cessed images captured through the mobile applica-

tion. We also introduced an incremental update pro-

cess that ensures the adaptability of YOLOv5-n to

real-life images collected through the mobile appli-

cation. Our experimental results showed the perfor-

mance of our deployed model in terms of classiﬁca-

tion accuracy, precision, recall, and mean average pre-

cision.

In the future, we intend to release our mobile ap-

plication on the Play Store and make it available for

free. This will allow us to collect annotated real-life

data. The collected data will be used to improve the

performance of our mobile application and effectively

assess and validate our incremental update process.

Moreover, the weights of our pre-trained YOLOv5-

n model can serve as initial weights when we extend

our task to include the detection of multiple leaves per

image and identifying their status.

ACKNOWLEDGEMENTS

This work is supported by the German Academic Ex-

change Service (DAAD) in the Ta’ziz Science Coop-

erations Program (AirFit Project; 57682841).

REFERENCES

Ahmed, A. A. and Reddy, G. H. (2021). A mobile-based

system for detecting plant leaf diseases using deep

learning. AgriEngineering, 3:478–493.

Aishwarya, N., Praveena, N., Priyanka, S., and Pramod, J.

(2023). Smart farming for detection and identiﬁca-

tion of tomato plant diseases using light weight deep

neural network. Multimedia Tools and Applications,

82:18799–18810.

Alshammari, H. H. and Alkhiri, H. (2023). Optimized re-

current neural network mechanism for olive leaf dis-

ease diagnosis based on wavelet transform. Alexan-

dria Engineering Journal, 78:149–161.

Alshammari, H. H., Taloba, A. I., and Shahin, O. R.

(2023). Identiﬁcation of olive leaf disease through

optimized deep learning approach. Alexandria Engi-

neering Journal, 72:213–224.

Canny, J. (1986). A computational approach to edge de-

tection. IEEE Transactions on Pattern Analysis and

Machine Intelligence, PAMI-8:679–698.

Chandan, J., Latha, D., Manisha, R., and Kishore, G.

(2022). Monitoring plant health and detection of plant

disease using iot and ml. International Journal of Re-

search in Engineering and Science, 10:1580–1588.

Chen, G., Choi, W., Yu, X., Han, T., and Chandraker, M.

(2017). Learning efﬁcient object detection models

with knowledge distillation. In Proceedings of Neural

Information Processing Systems (NIPS), pages 742–

751, Long Beach, CA, USA. Curran Associates, Inc.

Gou, J., Sun, L., Yu, B., Wan, S., and Tao, D. (2023). Hier-

archical multi-attention transfer for knowledge distil-

lation. ACM Transactions on Multimedia Computing,

Communications and Applications, 20:1–20.

Gurav, S. S., Patil, S. A., Patil, S. L., Desai, P. P., and Bad-

kar, P. B. (2022). Android application on e-commerce

& plant disease identiﬁcation using tensor ﬂow. Inter-

national Journal of Innovative Science and Research

Technology, 7:2456–2165.

Hidayat, A., Darusalam, U., and Irmawati, I. (2019). De-

tection of disease on corn plants using convolutional

neural network methods. Jurnal Ilmu Komputer dan

Informasi, 12:51–56.

Hinton, G., Vinyals, O., Dean, J., et al. (2015). Distill-

ing the knowledge in a neural network. In NIPS

Deep Learning and Representation Learning Work-

shop, Montreal, Canada. Curran Associates, Inc.

Kruthiventi, S. S., Sahay, P., and Biswal, R. (2017).

Low-light pedestrian detection from rgb images using

multi-modal knowledge distillation. In Proceedings

of the International Conference on Image Processing

(ICIP), pages 4207–4211, Beijing, China. IEEE.

Ksibi, A., Ayadi, M., Souﬁene, B. O., Jamjoom, M. M., and

Ullah, Z. (2022). Mobires-net: a hybrid deep learning

model for detecting and classifying olive leaf diseases.

Applied Sciences, 12:10278.

Li, Z., Xu, P., Chang, X., Yang, L., Zhang, Y., Yao, L., and

Chen, X. (2023). When object detection meets knowl-

edge distillation: A survey. IEEE Transactions on Pat-

tern Analysis and Machine Intelligence, 45:10555–

10579.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Pro-

ceedings of the 13th European Conference (ECCV),

pages 740–755, Zurich, Switzerland. Springer.

Mamdouh, N. and Khattab, A. (2022). Olive leaf dis-

ease identiﬁcation framework using inception v3 deep

learning. In Proceedings of the IEEE International

Conference on Design & Test of Integrated Micro &

Nano-Systems (DTS), pages 1–6, Cairo, Egypt. IEEE.

Ney, B., Bancal, M., Bancal, P., Bingham, I., Foulkes, J.,

Gouache, D., Paveley, N., and Smith, J. (2013). Crop

architecture and crop tolerance to fungal diseases and

insect herbivory. mechanisms to limit crop losses. Eu-

ropean Journal of Plant Pathology, 135:561–580.

Osco-Mamani, E. F. and Chaparro-Cruz, I. N. (2023).

Highly accurate deep learning model for olive leaf

disease classiﬁcation: A study in tacna-per

u. Inter-

national Journal of Advanced Computer Science and

Applications, 14:851–860.

Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O. R.,

and Jagersand, M. (2020). U2-net: Going deeper with

LIDL4Oliv: A Lightweight Incremental Deep Learning Model for Classifying Olive Diseases in Images

593

nested u-structure for salient object detection. Pattern

Recognition, 106:107404.

Savary, S., Ficke, A., Aubertot, J.-N., and Hollier, C.

(2012). Crop losses due to diseases and their impli-

cations for global food production losses and food se-

curity. Food Security, 4:519–537.

Soyyigit, S. and Yavuzaslan, K. (2018). Complex net-

work analysis of international olive oil market. Tarim

Ekonomisi Dergisi, 24:117–129.

Steen Lumholt and Guido van Rossum (2009). https://docs.

python.org/3/library/tkinter.html. Accessed: 2023-11-

19.

Tian, Y., Krishnan, D., and Isola, P. (2020). Contrastive

representation distillation. In Proceedings of the

International Conference on Learning Representa-

tions(ICLR), Addis Ababa, Ethiopia.

Tzutalin (2015). https://github.com/tzutalin/labelImg. Ac-

cessed: 2023-11-19.

Uguz,S. (2020). https://github.com/sinanuguz/CNN olive

dataset. Accessed: 2023-11-19.

Valdoria, J. C., Caballeo, A. R., Fernandez, B. I. D., and

Condino, J. M. M. (2019). idahon: An android based

terrestrial plant disease detection mobile application

through digital image processing using deep learning

neural network algorithm. In Proceedings of the 4th

International Conference on Information Technology

(InCIT), pages 94–98, Bangkok, Thailand. IEEE.

Wang, L., Lu, H., Wang, Y., Feng, M., Wang, D., Yin, B.,

and Ruan, X. (2017). Learning to detect salient ob-

jects with image-level supervision. In Proceedings of

the IEEE Conference on Computer Vision and Pattern

Recognition (CVPR), pages 136–145, Honolulu, HI,

USA. IEEE.

Wang, T., Yuan, L., Zhang, X., and Feng, J. (2019). Dis-

tilling object detectors with ﬁne-grained feature imi-

tation. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 4933–4942, California, USA. IEEE.

Yadikar, N., Ubul, K., et al. (2023). A review of knowledge

distillation in object detection. IEEE Access, 11.

Ye, Y., Zhu, X., Chen, H., Wang, F., and Wang, C. (2023).

Identiﬁcation of plant diseases based on yolov5. In In-

ternational Conference on Intelligent Computing and

Signal Processing (ICSP), pages 1731–1734, Xi’an,

China. IEEE.

ICAART 2024 - 16th International Conference on Agents and Artiﬁcial Intelligence

594