Comparative Study Between Object Detection Models, for Olive Fruit

Fly Identiﬁcation

Margarida Victoriano

1 a

, Lino Oliveira

1 b

and H

elder P. Oliveira

1,2 c

INESC TEC, FEUP Campus, Porto, Portugal

Faculty of Sciences, University of Porto, Porto, Portugal

Keywords:

Image Processing, Object Detection and Classiﬁcation, Insect Management.

Abstract:

Climate change is causing the emergence of new pest species and diseases, threatening economies, public

health, and food security. In Europe, olive groves are crucial for producing olive oil and table olives; however,

the presence of the olive fruit ﬂy (Bactrocera Oleae) poses a signiﬁcant threat, causing crop losses and ﬁnancial

hardship. Early disease and pest detection methods are crucial for addressing this issue. This work presents

a pioneering comparative performance study between two state-of-the-art object detection models, YOLOv5

and YOLOv8, for the detection of the olive fruit ﬂy from trap images, marking the ﬁrst-ever application of

these models in this context. The dataset was obtained by merging two existing datasets: the DIRT dataset,

collected in Greece, and the CIMO-IPB dataset, collected in Portugal. To increase its diversity and size, the

dataset was augmented, and then both models were ﬁne-tuned. A set of metrics were calculated, to assess

both models performance. Early detection techniques like these can be incorporated in electronic traps, to

effectively safeguard crops from the adverse impacts caused by climate change, ultimately ensuring food

security and sustainable agriculture.

1 INTRODUCTION

Climate change has considerable effects on the agri-

cultural industry, especially in terms of pests and dis-

eases. It is possible for pests and illnesses to thrive

and spread as temperature and precipitation patterns

change, increasing agricultural damage. Temperature

variations may also affect the behavior and life cycle

of pests, enabling them to spread more quickly and

colonize new areas, increasing the risk of agricultural

damage and making pest management more difﬁcult.

The timing of crop growth and the presence of natu-

ral pest predators can also be impacted by changes in

precipitation and temperature, and this can lead to a

mismatch between crop phenology and insect popula-

tions, which can exacerbate pest damage (Nazir et al.,

2018).

The expected effects of climate change on the dy-

namics of pest and diseases in agriculture underline

the signiﬁcance of effective pest control strategies that

take these factors into account. In order to reduce

crop losses and use safer control management strate-

https://orcid.org/0009-0003-1840-4636

https://orcid.org/0000-0003-1036-1072

https://orcid.org/0000-0002-6193-8540

gies that safeguard the crop, human health, and the

environment, early pest detection is essential.

Our research focuses on the olive fruit ﬂy (Bac-

trocera oleae), which is a signiﬁcant threat to olive

trees in the European Union, particularly in Spain and

Italy, which have the largest cultivation areas (Euro-

stat, 2019).

Monitoring the quantity of pests that have been

caught in a trap within a given time frame is one of

the crucial aspects of conﬁrming the occurrence of an

outbreak. Experts typically have to go to the ﬁelds

to visually inspect each trap manually to assess in-

festations, which can be a cumbersome and expen-

sive task, especially in large and dispersed orchards

(Shaked et al., 2017; Remboski et al., 2018; Hong

et al., 2020). To address this, new approaches have

been used to more effectively monitor this as well as

other insects, including the installation of electronic

traps supplied with cameras to record images of the

traps that can then be evaluated using classiﬁcation

and detection models (Uzun et al., 2022; Le et al.,

2021; Diller et al., 2022).

Due to the high percentage of EU cropland de-

voted to olive trees and the recurring outbreaks of

olive fruit ﬂies, monitoring and controlling pests in

458

Victoriano, M., Oliveira, L. and Oliveira, H.

Comparative Study Between Object Detection Models, for Olive Fruit Fly Identiﬁcation.

DOI: 10.5220/0012361800003660

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

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 2: VISAPP, pages

458-465

ISBN: 978-989-758-679-8; ISSN: 2184-4321

olive orchards is crucial to preserving the health and

productivity of olive trees as well as the quality of

by-products like table olives and olive oil. Early de-

tection of infestations and proactive approach through

control methods, like precise application of pesti-

cides, can reduce the need for more extensive inter-

ventions, assisting in the sustainable development of

olive production while conserving the environment

and human health (Caselli and Petacchi, 2021).

An approach that has been increasingly followed

is to use images collected in real time, from traps

placed on the ground, and computer vision algorithms

to automatically detect the presence of pests.

The purpose of this research is to provide a thor-

ough evaluation of the performance of two state-of-

the-art, fast, and accurate object identiﬁcation al-

gorithms, YOLOv5 and YOLOv8, in a novel area.

To accomplish this, we ﬁne-tuned the YOLOv5 and

YOLOv8 models using trap images collected from

two locations: Portugal and Greece. These computer

vision models combined with the prediction monitor-

ing process can be used to create a reliable and effec-

tive method for controlling olive fruit ﬂy populations

and protecting olive harvests.

2 RELATED WORK

Object detection is a crucial ﬁeld in computer vision

that deals with identifying and locating objects in im-

ages or videos. With the advent of deep learning tech-

niques, there has been substantial progress in object

detection in recent years. Insect detection and classi-

ﬁcation is a particular domain of object detection that

has signiﬁcant applications in agriculture, pest con-

trol, and disease prevention. In this section, we will

examine the current state-of-the-art in insect detection

and classiﬁcation, with a particular emphasis on deep

learning methods.

Deep learning is becoming increasingly signiﬁ-

cant for insect classiﬁcation and identiﬁcation due to

its capacity to automatically extract attributes from

raw visual data and learn intricate correlations be-

tween them. For a variety of purposes, including agri-

culture, pest management, and disease prevention, in-

sect classiﬁcation and identiﬁcation are essential. Ac-

curate identiﬁcation and classiﬁcation of insects can

be accomplished through the use of deep learning,

which can contribute in the creation of efﬁcient man-

agement and control techniques for insect populations

(Bjerge et al., 2022; Diller et al., 2022; Mamdouh and

Khattab, 2021; Tetila et al., 2020).

Diller et al. (Diller et al., 2022) conducted a study

to train deep learning algorithms for the identiﬁca-

tion of three major fruit ﬂy species - Ceratitis cap-

itata, Bactrocera dorsalis, and B. zonata. The al-

gorithm was trained on images collected from lab-

oratory colonies of two species and tested in the

ﬁeld in ﬁve different countries, and on images cap-

tured by an electronic trap built by the authors. The

electronic trap mostly contained only one specimen,

which made it easier to label each image and ap-

ply that label to each insect found in it. To ad-

dress the issue of multiple insect species in ﬁeld im-

ages, the authors manually labelled each occurrence

by drawing bounding boxes around them using spe-

ciﬁc software. Several data augmentation methods

were applied, such as horizontal ﬂipping, vertical ﬂip-

ping, rotations and random brightness and contrast

changes, to enhance the dataset and prevent overﬁt-

ting. The model was trained using the Faster R-CNN

ResNet50 algorithm and achieved an average preci-

sion of 87.41% to 95.80% for all three target classes.

The model also showed robustness to distinct lighting

conditions and accuracy on clustered samples. The

classiﬁcation accuracy for all classes in the dataset

collected on the ﬁeld was 98.29%, and the electronic

trap obtained high precision and accuracy for the three

target species, ranging from 86% to 97%. The pro-

posed model showed good ability to distinguish be-

tween the target classes and non-target classes.

Mamdouh and Khattab (Mamdouh and Khattab,

2021) proposed a framework to detect and count olive

fruit ﬂies in images captured by smart traps in olive

groves. The framework includes preprocessing of

the input dataset by normalizing it with yellow mean

color to unify the background color and make it il-

lumination invariant. The dataset images have vary-

ing lighting settings caused by varying light intensi-

ties and shadows cast by objects, so data augmenta-

tion techniques were applied (ﬂipping, shifting and

rotations) to increase the variability of images, thus

increasing accuracy. To reduce overﬁtting, the re-

searchers added a Leeds butterﬂy dataset to the train-

ing data. The researchers used YOLOv4, a single-

stage object detection paradigm, which requires only

one network pass to identify objects. Bounding box

detection is modelled as a regression task, and classes

are modelled as conditional probabilities. With a pre-

trained network, the ﬁnal layer is ﬁne-tuned based on

the input dataset. The network size was changed to

detect small objects, and the anchor boxes were clus-

tered based on the dataset bounding boxes dimen-

sions by running k-means with nine clusters. The

performance of the proposed framework was evalu-

ated using several metrics, including mean average

precision, precision, recall, and F1-score. The perfor-

mance metrics of the proposed framework were bet-

Comparative Study Between Object Detection Models, for Olive Fruit Fly Identiﬁcation

459

ter, comparing to the baseline model, although the

training time has greatly increased. The proposed

framework obtained 84% precision, 97% recall, 90%

F1-Score, and 96.68% mAP, while the baseline model

achieved 73% precision, 90% recall, 81% F1-Score,

and 87.59% mAP.

Tetila et al. (Tetila et al., 2020) proposed a frame-

work for classifying and counting insect pests in soy-

bean ﬁelds using images captured in the ﬁeld. The

authors evaluated three deep neural networks trained

with three different strategies: ﬁne-tuning of the net-

work layers with the weights obtained from Ima-

geNet, complete network with the weights initialized

randomly, and transfer learning with the weights ob-

tained from ImageNet. The proposed solution fol-

lows the SLIC Superpixels method to segment the

insects by employing the k-means algorithm to gen-

erate the superpixels that represent the similar re-

gions. Image annotation is performed by specialists

to compose a dataset for training and testing purposes.

The obtained dataset comprises seven insect classes.

A CNN is then trained to extract visual characteris-

tics from superpixels and produce the classiﬁcation

model to categorize the insect images. In the post-

processing stage, a plantation image is segmented us-

ing the SLIC method, and each segment’s superpixels

are assigned to a certain class. To identify each super-

pixel using the CNN-trained classiﬁcation model, the

system scans the image from left to right and top to

bottom while parallelly displaying the class’s color.

Thus, by supplying one class per segment, a color-

ful map is produced. The number of insects is de-

termined by adding up the insects from each super-

pixel class that the system has identiﬁed. The dataset

was augmented using rotation, scaling, scrolling, and

zooming. For image classiﬁcation, the authors trained

with Inception-Resnet-v2, ResNet50, and DenseNet-

201. The DenseNet-201 obtained the best accuracy,

94.89%, followed by Resnet-50 with 93.78% and then

Inception-Resnet-v2 with 93.40%.

Bjerge et al. (Bjerge et al., 2022) developed a real-

time insect detection and classiﬁcation algorithm us-

ing YOLOv3 deep learning model. The study aimed

to identify eight different insect species by exploring

various combinations of network parameters, such as

image size, kernel size, and number of training itera-

tions, to obtain the best model performance. The au-

thors acquired a dataset of 5757 images containing

insects and 2121 background images without insects

from Sedum plants and Calluna vulgaris plants in

Denmark, which were manually labelled with bound-

ing boxes surrounding the insects. The dataset was

augmented using default settings of the algorithm to

increase the size of the training dataset. The authors

evaluated the performance of the model using several

metrics, including precision and recall, during train-

ing and testing phases. The average training preci-

sion was 95%, and the average training recall was

91%. For testing, the average precision and recall

were 72% and 73%, respectively, and the mean av-

erage precision (mAP) was 87%. Overall, the study

demonstrated the ability of YOLOv3 to accurately de-

tect and classify different insect species in real-time.

In summary, while the reviewed studies demon-

strate the potential of deep learning algorithms for

insect classiﬁcation and detection, there are still op-

portunities for further improvement in terms of data

augmentation, labelling, pre-training techniques and

the use of different networks, namely YOLOv5 and

YOLOv8, especially for the detection of the olive fruit

ﬂy, due to the lack of comprehensive studies on this

particular insect.

3 PROPOSED METHODOLOGY

Our study introduces a comparison between two mod-

els that can identify the olive fruit ﬂy, also known as

Bactrocera oleae, a pest that has a signiﬁcant negative

impact on the olive oil industry. We used YOLOv5

(Jocher, 2020) and YOLOv8 (Jocher et al., 2023) ar-

chitectures for this purpose, due to their high accu-

racy shown on several object detection benchmarks,

fast inference and ﬂexibility. Our framework involves

multiple steps, including dataset preparation, applica-

tion of data augmentation techniques to mimic real-

world scenarios, and training with both models on a

cross-cohort dataset of olive fruit ﬂy trap images. To

assess the performance of both models, we used three-

fold cross-validation with hyperparameter tuning.

3.1 Dataset Description

The CIMO-IPB dataset (Pereira, 2023) contains 321

pictures, collected in Portuguese olive groves, mostly

exhibiting olive fruit ﬂies, trapped in yellow sticky

traps. Experts annotated each image in the dataset

with spatial identiﬁcation of the olive fruit ﬂy using

the labelImg tool (Tzutalin, 2015). The insects in the

images were manually outlined using bounding boxes

and labelled, as shown in Figure 1.

The DIRT dataset (Kalamatianos et al., 2018)

comprises 848 pictures, predominantly exhibiting

olive fruit ﬂies trapped in McPhail traps from 2015

to 2017 at different sites in Corfu, Greece. While the

majority of the images were procured via the e-Olive

mobile application, allowing users to upload captures

and photographs, they were obtained using a range of

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

460

Figure 1: CIMO-IPB Dataset Image Sample with Annota-

tions (Pereira, 2023).

gadgets like smartphones, tablets, and cameras, which

makes the dataset non-standardized. Each image in

the dataset contains spatial identiﬁcation of the olive

fruit ﬂy, annotated by experts, using the labelImg tool,

where the insects in the images were outlined manu-

ally using bounding boxes and tagged. Figure 2 shows

an annotated image retrieved from the DIRT dataset.

Figure 2: DIRT Dataset Image Sample with Annotations

(Kalamatianos et al., 2018).

Both datasets contain other insects, but those in-

stances will be considered as background.

We merged both image datasets, in order to add

more variability to the data. By combining these two

datasets, we were able to create a more comprehen-

sive and more diverse dataset that included images

of the same insect species but with different back-

grounds, lighting conditions, and poses, allowing the

model to generalize better to new, unseen data.

3.2 Data Preprocessing

In order to use the labels with the YOLOv5 and

YOLOv8 object detection model, the labels ﬁrst had

to be transformed from the PASCAL VOC XML for-

mat to the YOLO format. The annotations in the

YOLO format are described by a text ﬁle, for each

image, where each line represents an object. Each ob-

ject is described with an identiﬁer, followed by the

x-coordinate and the y-coordinate of the center of the

bounding box, and the width and height of the bound-

ing box. We followed a common practice of dividing

the dataset into training, validation, and testing sets,

where 80% was assigned for training, 10% for valida-

tion, and 10% for testing. This enabled us to employ

most of the data for model training while holding back

a section to evaluate its performance. Furthermore, to

enhance the model’s precision and robustness, we im-

plemented three-fold cross-validation, on the training

set, in addition to the split.

3.3 Data Augmentation Techniques

Performing object detection on trap images captured

outdoors may pose some challenges that may under-

mine the model’s accuracy and robustness. Some

of these difﬁculties, for instance, include variations

in lighting, weather conditions, and the presence of

blurry or low-quality images. To overcome this, data

augmentation is a crucial stage in the process of train-

ing models. The goal is to increase the variety and

volume of data that is available, which will boost

the model’s precision, resilience, and generalizabil-

ity. Figure 3 shows the data augmentation techniques

that were applied to the dataset.

We applied horizontal (Figure 3b) and vertical

ﬂips (Figure 3c), which is an efﬁcient way to create

mirrored versions of the original images. This in-

creases the quantity of training data that is available

and enables the model to learn to detect objects from

a variety of orientations.

We also applied blurring (Figure 3d), which repli-

cates ﬂuctuations in focus and camera movement that

may come from external factors. Similarly, random

brightness modiﬁcation (Figure 3g) can aid in simu-

lating changes in lighting circumstances, which can

be inﬂuenced by weather or sun position. We also ap-

plied multiplicative noise (Figure 3e and Figure 3f),

which adds noise to the images to help the model

learn to recognize insects under noisy conditions.

By using JPEG compression (Figure 3h), we com-

pressed the images to enable the model to learn to rec-

ognize insects under low-resolution conditions. This

is particularly useful in the context of capturing pic-

tures in the ﬁeld, where images need to be compressed

to reduce their size for transmission or storage, lead-

ing to poorer image quality.

3.4 YOLOv5 Architecture

YOLO (You Only Look once) (Redmon et al., 2016)

is a real-time object detection algorithm. YOLOv5

(Jocher, 2020) is the ﬁfth version of the YOLO algo-

rithm and is designed to be faster, more accurate, and

easier to use than its predecessors.

YOLOv5 is a single-stage object detector and is

composed of three main components: the backbone,

Comparative Study Between Object Detection Models, for Olive Fruit Fly Identiﬁcation

461

(a) Original image. (b) Horizontal Flip. (c) Vertical Flip. (d) Blurr.

(e) Multiplicative Noise. (f) Multiplicative Noise

with varying factor.

(g) Random Brightness

Contrast.

(h) JPEG Compression.

Figure 3: Data augmentation techniques applied to the olive fruit ﬂy image dataset. This helped to improve the performance

of the models to detect olive fruit ﬂies in different orientations and in different conditions.

the neck and the head.

Yolov5 combined the cross stage partial network

(CSPNet) (Wang et al., 2019) with Darknet to create

CSPDarknet, which operated as the network’s back-

bone. By including gradient changes into the fea-

ture map, this integration decreases the parameters

and FLOPS of the model and addresses the issue of

redundant gradient information in large-scale back-

bones. This not only guarantees quicker and more

precise inference, but also minimizes the dimension

of the whole model.

In order to improve information ﬂow, Yolov5 ad-

ditionally uses the path aggregation network (PANet)

(Wang et al., 2020) as its neck. The dissemination

of low-level features is enhanced by PANet’s novel

feature pyramid network (FPN) structure, which

strengthens the bottom-up path. Moreover, adaptive

feature pooling is used to link feature grids and all fea-

ture levels, enabling helpful information at each level

to spread immediately to the next subnetwork.

The model uses SPPF, which is a variant of Spatial

Pyramid Pooling (SPP) (He et al., 2014) that com-

bines data from inputs and produces a ﬁxed-length

output, without slowing down the network. The most

important context features are separated, and the re-

ceptive ﬁeld is greatly expanded.

Yolov5’s head (Redmon and Farhadi, 2018) also

produces feature maps in three distinct sizes to pro-

vide multi-scale prediction and aid in the detection of

small, medium, and large objects. It predicts the ob-

ject bounding boxes, the scores and the object classes.

YOLOv5 architecture overview is described in

Figure 4.

Figure 4: YOLOv5 Architecture.

3.5 YOLOv8 Architecture

YOLOv8 (Jocher et al., 2023) is the successor of the

YOLOv5 algorithm and presents an integrated frame-

work for training models for object detection, in-

stance segmentation and image classiﬁcation.

YOLOv8 introduces a new backbone, Darknet-53,

that is faster and more accurate than previous ver-

sions. The effectiveness of this version is due to the

use of a larger feature map and a more effective con-

volutional network, which leads to faster and more ac-

curate object detection. The model’s capacity to iden-

tify patterns and objects in the data is improved by the

larger feature map’s ability to support more complex

feature interactions. It also helps to lessen overﬁtting

and model training time.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

462

Moreover, YOLOv8 makes use of feature pyramid

networks, which boosts overall accuracy by better

identifying objects of various sizes. Feature Pyramid

networks use different feature map scales in conjunc-

tion with skip connections to more accurately predict

larger and smaller objects, analogous to making pre-

dictions on different image sizes.

Unlike the YOLOv5, the YOLOv8 introduced an

anchor-free detection head. It does not use anchor

boxes to predict the offset from a predeﬁned anchor

box. Instead, it directly predicts the object’s center in

an image.

Figure 5 illustrates YOLOv8 architecture.

Figure 5: YOLOv8 Architecture.

3.6 Training and Testing

We conducted three-fold cross-validation with hyper-

parameter tuning on the YOLOv5 and YOLOv8 mod-

els using previously split training data. This allowed

us to assess the performance of both models on differ-

ent subsets of the data and optimize the hyperparam-

eters.

The hyperparameters we experimented with are

summarized on the Table 1.

Table 1: Range of hyperparameters varied during experi-

mentation.

Hyperparameter Range

Number of Epochs 50 - 200

Batch Size 8 - 20

Learning Rate 0.0001 - 0.1

Weight Decay 0.0005 - 0.001

Optimizers SGD and Adam

4 RESULTS EVALUATION

Precision, recall, and mean average precision were

used to evaluate the models’ performance after cross-

validation.

We retrained both models using the best perform-

ing hyperparameters, obtained with cross-validation,

as shown in Table 2.

Table 2: Best performing hyperparameters in YOLOv5 and

YOLOv8.

Model Number of Epochs Batch Size Learning Rate Weight Decay Optimizer

YOLOv5 150 8 0.01 0.0005 SGD

YOLOv8 90 8 0.01 0.0005 SGD

Finally, we tested the retrained models on the 10%

previously split set.

The test results show that YOLOv5 and YOLOv8

perform similarly under identical training conditions.

YOLOv5 achieved a precision of 0.8, recall of 0.75,

and mAP of 0.82 at an IoU threshold of 0.5 dur-

ing cross-validation. After retraining, its precision

slightly dropped to 0.76, while recall and mAP re-

mained constant.

On the other hand, YOLOv8, during cross-

validation, attained a precision of 0.75, recall of 0.73,

and mAP of 0.79. After retraining, its precision im-

proved to 0.79, with recall staying at 0.74 and mAP

increasing to 0.82.

In summary, both models performed well in in-

sect detection tasks. YOLOv5 demonstrated higher

precision during cross-validation, whereas YOLOv8

showed signiﬁcant improvement in precision and

mAP after retraining. Notably, YOLOv8 trained

faster than YOLOv5, which is crucial for efﬁcient re-

source utilization in object detection tasks.

Table 3 shows the comparison between the ob-

tained results.

Table 3: Results comparison between YOLOv5 and

YOLOv8.

Metric

3-fold Cross Validation Testing with Best Parameters

YOLOv5 YOLOv8 YOLOv5 YOLOv8

Precision 0.80 0.75 0.76 0.79

Recall 0.75 0.73 0.75 0.74

mAP@.5 0.82 0.79 0.82 0.82

Figure 6 shows the detection results obtained us-

ing YOLOv5 and YOLOv8 models, respectively, on

samples from both datasets. On the DIRT dataset,

both models detected correctly the olive fruit ﬂy in-

stances; however, YOLOv5 predictions had higher

conﬁdence values (Figure 6a), in comparison to

YOLOv8 conﬁdence values (Figure 6c). On the other

hand, on the CIMO-IPB dataset, both models per-

formed similarly in terms of the detection conﬁdence

values, as seen in Figure 6b and Figure 6d. The

Comparative Study Between Object Detection Models, for Olive Fruit Fly Identiﬁcation

463

(a) Detection results on

DIRT dataset sample us-

ing YOLOv5.

(b) Detection results on

CIMO-IPB dataset sam-

ple using YOLOv5.

DIRT dataset sample us-

ing YOLOv8.

(d) Detection results on

CIMO-IPB dataset sam-

ple using YOLOv8.

Figure 6: Detection results using YOLOv5 and YOLOv8.

CIMO-IPB dataset posed a set of challenges. Many

images were marred by shadows, cast by natural el-

ements or artiﬁcial structures (like the phone shadow

in Figure 6b), which can obscure crucial details. Ad-

ditionally, intense sunlight created areas of overex-

posure, making it difﬁcult to discern ﬁner features.

Furthermore, a lack of precise focus in some images

compounded the complexity of the dataset. Despite

the lower scores on these images, the models were

still able to correctly predict the target class. This is a

positive result, as it indicates that the models are ro-

bust to real-world challenges such as shadows, intense

sun, and lack of focus. In real-world applications, it is

important to have models that can still perform well

even when faced with these challenges.

In the literature review, it is noted that a method

proposed by (Mamdouh and Khattab, 2021) employ-

ing the DIRT dataset yielded superior results com-

pared to our own approach. While both methodolo-

gies share the foundation of the DIRT dataset, a crit-

ical distinction arises from the integration of an ad-

ditional dataset, the CIMO-IPB dataset, in our work.

The CIMO-IPB dataset encompasses images of the

same target object, the olive fruit ﬂy, yet the con-

ditions and settings of capture signiﬁcantly diverge

from those in the DIRT dataset. CIMO-IPB dataset

comprises images of the same target object, albeit

captured under vastly different conditions and set-

tings. This strategic integration aims to enhance the

robustness and generalizability of our model, a fac-

tor not accounted for in the comparison. The rea-

soning behind combining the CIMO-IPB dataset and

the DIRT dataset was to increase the diversity and

robustness of our model. By combining images col-

lected under different situations, our method develops

a more comprehensive grasp of the olive fruit ﬂy’s de-

piction. In comparison to the method that only uses

the DIRT dataset, our mixed dataset approach has sig-

niﬁcant advantages in terms of improved generaliz-

ability and adaptability across varied settings. The

suggested approach by Mamdouh and Khattab (Mam-

douh and Khattab, 2021) achieve a slightly higher

precision than our method, 84% versus 79%, while

it surpasses it on the other metrics, due to the speciﬁc

characteristics of the DIRT dataset.

In summary, the alternative technique that focuses

entirely on the DIRT dataset produces great results,

particularly in terms of recall and mAP. This is most

likely owing to the dataset’s narrow focus. Our tech-

nique, on the other hand, which mixes two datasets

with differing conditions, maintains competitive per-

formance across all criteria. This shows that it has

the ability to perform well in a variety of real-world

circumstances. The decision to mix the DIRT dataset

and the CIMO-IPB dataset exempliﬁes an intentional

effort to improve the model’s adaptability and versa-

tility, as evidenced by the acceptable results obtained.

During our research, we noticed that there is a lack

of publicly available datasets containing the olive fruit

ﬂy. Due to the limited availability of datasets for our

target species, we explored related studies on differ-

ent insects, to potentially adapt methodologies from

these works. Comparing the results between studies

is not signiﬁcant, due to their different morphologies

and features. Our work was impacted by the lack

of comprehensive datasets, but our tailored approach

was shown to be effective for the detection of the olive

fruit ﬂy.

5 CONCLUSION AND FUTURE

WORK

To summarize, the purpose of this study was to assess

two alternative approaches for identifying and catego-

rizing olive fruit ﬂies, which pose a signiﬁcant risk to

olive cultivation in the European Union. The method-

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

464

ology consisted of augmenting the dataset and ﬁne-

tuning YOLOv5 and YOLOv8 models using the aug-

mented data. The experimental results showed that

both models effectively detected the olive fruit ﬂy,

but YOLOv8 obtained superior results, in terms of the

evaluation metrics.

In the future, the ﬁne-tuned models, which include

domain knowledge to help improve the methodolo-

gies, are going to be integrated on a web-based infor-

mation and management system, which will receive

the images collected from the ﬁeld, by a robotic smart

trap developed by our team, and return the image

with the detection results, to aid farmers in making

informed decisions.

ACKNOWLEDGEMENTS

This work is co-ﬁnanced by Component 5 - Capital-

ization and Business Innovation of core funding for

Technology and Innovation Centres (CTI), integrated

in the Resilience Dimension of the Recovery and Re-

silience Plan within the scope of the Recovery and

Resilience Mechanism (MRR) of the European Union

(EU), framed in the Next Generation EU, for the pe-

riod 2021 - 2026

REFERENCES

Bjerge, K., Mann, H. M., and Høye, T. T. (2022). Real-time

insect tracking and monitoring with computer vision

and deep learning. Remote Sensing in Ecology and

Conservation, 8:315–327.

Caselli, A. and Petacchi, R. (2021). Climate change and

major pests of mediterranean olive orchards: Are we

ready to face the global heating? Insects, 12(9).

Diller, Y., Shamsian, A., Shaked, B., Altman, Y., Danziger,

B. C., Manrakhan, A., Serfontein, L., Bali, E., Wer-

nicke, M., Egartner, A., Colacci, M., Sciarretta, A.,

Chechik, G., Alchanatis, V., Papadopoulos, N. T., and

Nestel, D. (2022). A real-time remote surveillance

system for fruit ﬂies of economic importance: sen-

sitivity and image analysis. Journal of Pest Science,

1:1–12.

Eurostat, E. (2019). Statistics explained. (Accessed on

2/02/2023).

He, K., Zhang, X., Ren, S., and Sun, J. (2014). Spatial

pyramid pooling in deep convolutional networks for

visual recognition. In Fleet, D., Pajdla, T., Schiele, B.,

and Tuytelaars, T., editors, Computer Vision – ECCV

2014, pages 346–361, Cham. Springer International

Publishing.

Hong, S.-J., Kim, S.-Y., Kim, E., Lee, C.-H., Lee, J.-S.,

Lee, D.-S., Bang, J., and Kim, G. (2020). Moth detec-

tion from pheromone trap images using deep learning

object detectors. Agriculture, 10(5):170.

Jocher, G. (2020). YOLOv5 by ultralytics. Version 7.0.

Jocher, G., Chaurasia, A., and Qiu, J. (2023). YOLO by

Ultralytics.

Kalamatianos, R., Karydis, I., Doukakis, D., and Avlonitis,

M. (2018). Dirt: The dacus image recognition toolkit.

Journal of Imaging 2018, Vol. 4, Page 129, 4:129.

Le, A. D., Pham, D. A., Pham, D. T., and Vo, H. B. (2021).

Alerttrap: A study on object detection in remote in-

sects trap monitoring system using on-the-edge deep

learning platform.

Mamdouh, N. and Khattab, A. (2021). Yolo-based deep

learning framework for olive fruit ﬂy detection and

counting. IEEE Access, 9.

Nazir, N., Bilal, S., Bhat, K., Shah, T., Badri, Z., Bhat,

F., Wani, T., Mugal, M., Parveen, S., and Dorjey, S.

(2018). Effect of climate change on plant diseases. In-

ternational Journal of Current Microbiology and Ap-

plied Sciences, 7:250–256.

Pereira, J. A. (2023). Yellow sticky traps dataset

olive ﬂy

(Bactrocera Oleae).

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

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

ject detection.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement.

Remboski, T. B., de Souza, W. D., de Aguiar, M. S., and

Ferreira, P. R. (2018). Identiﬁcation of fruit ﬂy in in-

telligent traps using techniques of digital image pro-

cessing and machine learning. In Proceedings of the

33rd Annual ACM Symposium on Applied Computing,

SAC ’18, page 260–267, New York, NY, USA. Asso-

ciation for Computing Machinery.

Shaked, B., Amore, A., Ioannou, C., Vald

es, F., Alorda,

B., Papanastasiou, S., Goldshtein, E., Shenderey, C.,

Leza, M., Pontikakos, C., and et al. (2017). Electronic

traps for detection and population monitoring of adult

fruit ﬂies (diptera: Tephritidae). Journal of Applied

Entomology, 142(1-2):43–51.

Tetila, E. C., MacHado, B. B., Menezes, G. V., Belete, N.

A. D. S., Astolﬁ, G., and Pistori, H. (2020). A deep-

learning approach for automatic counting of soybean

insect pests. IEEE Geoscience and Remote Sensing

Letters, 17:1837–1841.

Tzutalin (2015). Labelimg.

Uzun, Y., Tolun, M. R., Eyyuboglu, H. T., and Sari, F.

(2022). An intelligent system for detecting mediter-

ranean fruit ﬂy [medﬂy; ceratitis capitata (wiede-

mann)]. Journal of Agricultural Engineering, 53.

Wang, C.-Y., Liao, H.-Y. M., Yeh, I.-H., Wu, Y.-H., Chen,

P.-Y., and Hsieh, J.-W. (2019). Cspnet: A new back-

bone that can enhance learning capability of cnn.

Wang, K., Liew, J. H., Zou, Y., Zhou, D., and Feng, J.

(2020). Panet: Few-shot image semantic segmenta-

tion with prototype alignment.

Comparative Study Between Object Detection Models, for Olive Fruit Fly Identiﬁcation

465