IDiSSC: Edge-computing-based Intelligent Diagnosis Support System

for Citrus Inspection

Mateus Coelho Silva

1,2 a

, Jonathan Crist

ao Ferreira da Silva

1 b

and Ricardo Augusto Rabelo Oliveira

2 c

Instituto Federal de Educac¸

ao, Ci

encia e Tecnologia de Minas Gerais, Campus Avanc¸ado Itabirito, Brazil

Departamento de Computac¸

ao, Instituto de Ci

encias Exatas e Biol

ogicas, Universidade Federal de Ouro Preto, Brazil

Keywords:

Edge Computing, Embedded Systems, Computer Vision, Machine Learning.

Abstract:

Orange and citrus agriculture has a signiﬁcant economic role, especially in tropical countries. The use of

edge systems with machine learning techniques presents a perspective to improve the present techniques, with

faster tools aiding the inspection diagnostics. The usage of cost- and resource-restrictive devices to create

these solutions improves this technique’s reach capability and reproducibility. In this perspective, we propose

a novel edge-computing-based intelligent diagnosis support system performing a pseudospectral analysis to

improve the orange inspection processes. Our results indicate that traditional machine learning methods reach

over 92% accuracy, reaching 99% on the best performance technique with Artiﬁcial Neural Networks in the

binary classiﬁcation stage. For multiple classes, the accuracy varies from 97% up to 98%, also reaching the

best performance with Artiﬁcial Neural Networks. Finally, the Random Forest and Artiﬁcial Neural Network

obtained the best results, considering algorithm parameters and embedded hardware performance. These

results enforce the feasibility of the proposed application.

1 INTRODUCTION

Computer vision is increasingly being inserted in pro-

duction systems, revolutionizing the quality manage-

ment in the industry (4vision, 2019). A sector of ex-

treme global importance is agribusiness. This sector

can beneﬁt from these techniques for more modern,

economic, and safe processes. In this context, a com-

puter vision associated with edge systems becomes a

tool that enables technological advances in the ﬁeld,

such as in citrus. These techniques assist in develop-

ing the industry and contribute to the optimization of

traditional sectors of the economy (da Rosa, 2019).

However, the development of an intelligent algorithm

allows identifying diseases in oranges on a large scale

(Soini et al., 2019).

The National Association of Citrus Juice Ex-

porters in Brazil explains that orange is one of the

most cultivated fruits. Also, the fruit has a substan-

tial impact on the Brazilian economy. The cultiva-

tion of citrus fruits requires a large number of work-

https://orcid.org/0000-0003-3717-1906

https://orcid.org/0000-0003-2214-397X

https://orcid.org/0000-0001-5167-1523

ers and generates a GDP of US$ 6.5 billion in all

countries of the production chain (Neves and Trom-

bin, 2017). Brazilian orange juice exports grew by

26.6% between July and December 2019, taking into

account previous periods. This year’s volume went

from 512,388 tons to 648,751 tons and a turnover of

US$ 967.1 million to US$ 1,104 billion (CitrusBr,

2020).

Figure 1: An example of black spot disease in an orange.

Source: (Fundecitrus, 2019).

An important aspect of improving productivity in or-

ange crops is the detection of diseases through in-

spection. Among the major diseases in orange farms,

some of the main plagues are the black spot, cit-

Silva, M., Ferreira da Silva, J. and Oliveira, R.

IDiSSC: Edge-computing-based Intelligent Diagnosis Support System for Citrus Inspection.

DOI: 10.5220/0010444106850692

In Proceedings of the 23rd International Conference on Enterprise Information Systems (ICEIS 2021) - Volume 1, pages 685-692

ISBN: 978-989-758-509-8

685

rus canker, and Greening. These diseases’ impacts

are related to the reduction of product quality due to

fungi or bacteria, fruit damage, and production reduc-

tion due to the premature fall of fruits from the trees.

Also, commercialization becomes restricted due to

these factors that hinder production (Gottwald et al.,

2002). Another issue related to this disease is fruits

with acidic and bitter ﬂavors and a poor appearance

on the surface. Thus, these products become unsuit-

able for sales, such as fresh fruit and juice production

(USDA APHIS, 2018). Figure 1 displays an orange

with black spot fungus contamination.

In this work, we present IDiSSC: a novel system

to aid in the diagnostics of oranges’ diseases. The

method targets its usage on constrained edge comput-

ing devices. For this matter, we perform a study of

the classiﬁer candidates and evaluate the algorithms’

performance in different hardware conﬁgurations. Fi-

nally, we propose a system with an embedded com-

puter vision algorithm to classify the oranges and sug-

gest a diagnosis. Thus, the main contribution of this

work is:

• The proposal and proof-of-concept of a novel or-

ange disease detection diagnosis support system;

For evaluating this matter, Section 2 presents the the-

oretical references used in this work. In Section 3, we

describe the classiﬁcation system’s main features for

fresh and rotten oranges. We propose the methodol-

ogy for validating aspects of this system in Section 4.

In Section 5, we display the results of an analysis of

the data and its interpretation. Finally, in Section 6,

we discuss the results obtained and a comprehensive

discussion of this work.

2 BACKGROUND

In the previous section, we presented the motivation

and main general aspects of this work. In this section,

we present some machine techniques used for image

classiﬁcation. Also, we present some of the most rel-

evant related work and how they approach and differ

from our work.

2.1 Image Classiﬁcation using Machine

Learning

In this work, we consider some of the supervised ma-

chine learning techniques used to classify data. It is

essential to carry out the analysis of these classiﬁers

for the development of this work. For this matter, we

experimented with four different techniques of super-

vised machine learning:

• Structured Vector Machine (SVM)

• K-Nearest Neighbors (KNN)

• Random Forest

• Artiﬁcial Neural Network (ANN)

The SVM is a classiﬁer that separates classes by

creating a hyperplane, a separation line between the

two data groups. In the literature, some works use

SVM classiﬁers to detect leaf diseases (Padol and Ya-

dav, 2016), pest detection on strawberry greenhouses

(Ebrahimi et al., 2017), tumor detection no MRI im-

ages (Mathew and Anto, 2017), among others. As

in this work, these techniques also employ a previous

feature extraction stage.

Another method to classifying images is with the

K-Nearest Neighbors (KNN) classiﬁer. On the liter-

ature, authors used this technique to evaluate hyper-

spectral images (Huang et al., 2016; Tu et al., 2018),

MRI images (Wasule and Sonar, 2017), varicose ul-

cer detection (Bhavani and Wiselin Jiji, 2018), among

others. These works also employ the same steps, with

a feature extraction stage, followed by the classiﬁca-

tion using a machine learning technique.

Another important use of machine learning tech-

niques is artiﬁcial neural networks, which mimic

neurons’ functioning in the human nervous system.

Multi-Layer Perceptron (MLP) is a technique used

for image classiﬁcation. Some works use this tech-

nique, such as to classify images of mammography

exam (Valarmathie1 et al., 2016), blader cystoscopic

images classiﬁcation (Hashemi et al., 2020), diabetic

retinopathy fundus image classiﬁcation (Shankar

et al., 2020), and many others. The deep learning neu-

ral networks can avoid the stage of feature extraction

(Silva and Siebra, 2017; C¸ alik and Demirci, 2018).

Nonetheless, these techniques still have strong hard-

ware requirements (LeCun, 2019), which is not ideal

when considering cost- and resource-restrictive sys-

tems such as edge devices.

2.2 Oranges Classiﬁcation using Edge

Systems

In this section, we explore the works related to our

proposition, exposing similar and different features.

Rotondo et al. (Rotondo et al., 2018) present a system

based on an Android application and a cloud service.

It classiﬁes the acquired image among the produced

species, using a KNN classiﬁer with the Bag of Vi-

sual Words (BoW) as a feature vector. The method is

cloud-dependant, while our proposal realizes the pro-

cessing on the edge device.

Yin et al. (Yin et al., 2017) display a method for

identifying decayed oranges infected by fungi using

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

686

hyperspectral images. They do not use any machine

learning method. Initially, their system requires more

acquisition equipment to perform classiﬁcations, as

they provide a segmentation method based on hyper-

spectral images.

Putra et al. (Putra et al., 2018) developed a fea-

ture extraction system for modeling the quality of or-

anges. This system aims to support the classiﬁcation

of a sorting system. Thus, the authors mainly explore

the feature extraction process. It also performs image

segmentation, while in this work, our result does not

need to perform this extra task.

The previous works do not present a resource-

restrictive device to perform the same task as de-

scribed in this work. Also, the authors do not com-

pare the performance with different classiﬁers to un-

derstand how the algorithm’s change affects the clas-

siﬁcation performance. Finally, they do not consider

the timing restrictions for each algorithm when func-

tioning in edge devices.

3 SYSTEM DESCRIPTION

Figure 2: Portable classiﬁcation system.

In this section, we present the proposed system ar-

chitecture and its elements. Bearing in mind that

an agricultural technician needs to gather information

for analysis in the open ﬁeld, we propose an orange

classiﬁcation system, displayed in Figure 2. This ap-

pliance can provide the necessary information to the

technician in a short time, with greater precision. The

collected information can be sent via WLAN or stored

in a database for quality sectors.

3.1 Classiﬁcation Algorithm

The core feature of the detection system is the clas-

siﬁcation algorithm based on machine learning. We

propose to use computer vision and machine learn-

ing to classify oranges (Blasco et al., 2016). After

these stages, we expect to identify healthy and fresh

oranges from the image analysis. Figure 3 displays a

schematic view of the proposed classiﬁcation method

for fresh or rotten oranges. In the second stage of the

test, we also validated the classiﬁcation stage for mul-

tiple classes using the same proposed technique.

Figure 3: Computer-Vision-based machine learning algo-

rithm creation process.

These classiﬁers use computer vision to process these

fruits. We propose to use a “pseudospectral analysis”

(Puchkov and McCarren, 2011; Puchkov et al., 2016;

Pipatnoraseth et al., 2019) on the provided images.

For this matter, we initially take the original images

and convert them to the HSV color space. From this

information, we extract the color conﬁguration fea-

ture from the Hue channel, using it as a “pseudospec-

trum” of the composing colors. As no color space

represents the complete color spectrum from the ac-

tual acquired data, we name this a “pseudospectral

analysis”.

We then used the obtained pseudospectrum as a

feature vector to train the machine learning algo-

rithm, labeling the data in fresh and rotten classes.

The low cost of implementing the system can be ad-

vantageous compared to other existing ones, such as

the Prism-based multi-spectral cameras that empower

high-speed fruit sorting (JAI, 2020).

4 EXPERIMENTAL

METHODOLOGY

In this section, we present the experimental method-

ology to validate some aspects of the proposed solu-

tion. For this matter, we test two aspects of this sys-

tem: Algorithm classiﬁcation performance compar-

IDiSSC: Edge-computing-based Intelligent Diagnosis Support System for Citrus Inspection

687

ing multiple machine learning techniques and the per-

formance of all trained algorithms in embedded hard-

ware, compared to general-purpose personal comput-

ers with various hardware conﬁgurations.

Edge systems intended for machine learning tasks

that use image extraction must have at least a CPU

and RAM and a connected camera to capture im-

ages. This task requires processing power from the

hardware as it has a signiﬁcant memory expenditure.

Thus, the higher this processing capacity, the bet-

ter the system’s performance for graphical analysis

(Rong et al., 2017). A system option to perform these

simple tasks with low data volume and use of rasp-

berry pi 3B, as it has some the necessary resources

to handle image processing with low cost and energy

efﬁciency (Jaskolka et al., 2019).

4.1 Machine Learning Algorithm:

Development and Performance Tests

We developed different versions of the classiﬁcation

algorithm. Initially, we performed the image prepro-

cessing using the OpenCV framework. In this stage,

we opened the image, converted it to the HSV color

space, extracted the histogram from the Hue chan-

nel, and used it as the “pseudospectrum” feature vec-

tor. For this matter, we used the “OpenCV Library”.

Then, we used the python library “scikit-learn” na-

tive tools to train and produce four supervised ma-

chine learning models: an SVM classiﬁer, a KNN

classiﬁer, a Random Forest classiﬁer, and an MLP

(ANN) classiﬁer. Through experiments, we came to

conﬁgurations where the KNN classiﬁer used seven

neighbors, the Random Forest used four estimators

with a maximum depth of ten, and the ANN has four

layers of 32 neurons each.

As stated, the ﬁrst tested aspect of this system is

the classiﬁcation algorithm performance. For this pur-

pose, we trained our algorithm using a dataset with a

large number of single orange images (Kalluri, 2018).

This set contains separate folders for training and test-

ing. The training set contains 1466 images of fresh or-

anges and 1595 images from rotten images, with 3061

images. From this set, we randomly separated around

10% for validating the algorithm after the training

stage. The remaining 90% were used to train the al-

gorithm. The test set has 388 images from fresh or-

anges and 403 images from rotten oranges. For the

multiple class stage, we also used a Greening dataset

(Rauf et al., 2009) containing 1845 images and com-

plemented it with self-made photos of oranges with

black spot (1030) and canker (1002), using a white

light source and white paper as the background.

An essential part of evaluating methods is the

choice of metrics. There is an inﬂuence of each

metric applied to the machine learning performance.

Therefore, there may be discrepancies in comparing

values between classiﬁers (Kumar, 2017). We used

three standard metrics for machine learning evalua-

tion: Precision, Recall, and F1-Score.

Precision =

T P

T P + FP

(1)

Recall =

T P

T P + FN

(2)

F1-Score = 2 ×

Precision × Recall

Precision + Recall

(3)

In this context, true positive (T P) shows the samples

correctly obtained by the classiﬁer is by the positive

class, and the true negative (T N) represents the same

with the negative class. The false positive (FP) refers

to the classiﬁer’s result in which the model is incor-

rectly classiﬁed to the positive class and false negative

(FN) to the negative class incorrectly. Also, we used

the confusion matrix in the test dataset as a ﬁnal eval-

uation metric. This matrix displays the distribution of

correct and incorrect classiﬁcations for each class.

4.2 Embedded Hardware Performance

Test

In this stage, we tested the algorithm performance

on multiple devices for this issue, using PMML as a

model persistence framework (Guazzelli et al., 2009).

For this implementation, we used the Nyoka

library

for model persistence and pypmml

to execute the test

appliance.

The repetitive algorithm path has three main

stages: (i) loading the image to the memory, (ii) ex-

tracting the image feature vector, and (iii) classifying

the image submitting the feature vector to the model.

To understand the hardware constraints when execut-

ing the proposed methods, we measured the average

time required to perform each stage. We performed

this test in three different hardware options:

• A Desktop Personal Computer, with a 9th gener-

ation i5 processor, RTX 2060 super GPU, 32GB

of RAM, with the code and data stored in an SSD.

This machine runs a Debian-based Linux Opera-

tional System.

https://github.com/nyoka-pmml/nyoka

https://pypi.org/project/pypmml/

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

688

• A Laptop Personal Computer, with an 8th gen-

eration i5 processor, onboard Intel GPU, 8GB of

RAM, with the code and data stored in an HDD.

This machine runs a Debian-based Linux Opera-

tional System.

• An Edge Computing Device, with a Quad-Core

ARMv8 processor, onboard GPU, 1GB of RAM,

with the code and data stored in an SD card. This

computer runs a standard Debian-based Linux

distribution designed for this platform.

For each system, we evaluated every produced model.

We analyzed the average time to perform each of the

three tasks and used it to estimate how many predic-

tions the system can perform per second using each

solution. For each model, we executed the experi-

ment in the whole test section of the dataset. This

subset contains 403 images from rotten oranges and

388 images of fresh oranges, totalizing 791 runs.

5 RESULTS

In this section, we present the results obtained from

the proposed tests. Also, we display our preliminary

conclusions based on each result.

5.1 Machine Learning Algorithms:

Performance Tests

The ﬁrst employed tests are the machine learning met-

ric evaluation for the candidate models. As stated in

Section 4, we evaluate three metrics: Precision, Re-

call, and F1-Score. We also present the overall model

accuracy, which is the correct prediction ratio given

the whole prediction set. Table 1 displays the results

for the machine learning training process. All candi-

date algorithms presented an overall accuracy above

92%.

Table 1: Test Model - Binary Classiﬁcation.

Precision Recall F1-Score Support

SVM Results:

Rotten 0.96 0.89 0.92 151

Fresh 0.90 0.96 0.93 156

Accuracy = 92.8% 307

KNN Results:

Rotten 0.95 0.99 0.97 163

Fresh 0.99 0.94 0.96 144

Accuracy = 96.7% 307

Random Forest Results:

Rotten 0.96 0.98 0.97 140

Fresh 0.98 0.96 0.97 167

Accuracy = 97.1% 307

MLP (ANN) Results:

Rotten 0.99 0.99 0.99 162

Fresh 0.99 0.99 0.99 145

Accuracy = 99.0% 307

The SVM had the worst performance with an ac-

curacy of 92.8%. The detection of rotten oranges had

a precision of 0.96, with a recall of 0.89. This result

indicates that this algorithm has a lower performance

in detecting all the samples of rotten oranges. The

F1-Score, in this case, was 0.92. The detection of

fresh oranges had a precision of 0.90, recall of 0.96,

and F1-Score of 0.93. These results indicate that the

system has a better performance when ﬁnding all true

positive samples but has a higher false-positive clas-

siﬁcation rate. In both cases, the F1-Score indicates

the same quality displayed by general accuracy.

The third best algorithm was the KNN, with an

accuracy of 96.7%. For the rotten oranges, this al-

gorithm presented a precision of 0.95, recall of 0.99,

and F1-Score of 0.97. This result indicates that the

system detects most samples from this class but in-

cludes a high amount of false-positive samples. For

the fresh oranges, the precision was 0.99, the recall

was 0.94, and the F1-score was 0.96. This result in-

dicates that this system has a small number of false-

positive samples but misses many samples that should

be classiﬁed as positive. Also, in both cases, the F1-

Score indicates the same quality displayed by general

accuracy.

The second-best performance was achieved by the

Random Forest model, with a general accuracy of

97.1%. For the rotten oranges, we observed preci-

sion of 0.96, recall of 0.98, and F1-Score of 0.97. For

the fresh oranges, the system presented a precision of

0.98, recall of 0.96, and F1-Score of 0.97. This model

is well balanced, detecting more fake rotten oranges

than fake fresh. The F1-Score corroborates these pre-

liminary conclusions.

The MLP model achieved the best performance on

this test. Its general average was 99.0%. For both

fresh and rotten orange sets, the precision, recall, and

F1-scores were 0.99. Thus, this model presented the

highest precision and balance considering all the sam-

ples.

After evaluating the validation set, we also per-

formed an experiment on the test dataset. We eval-

uated the confusion matrix for each classiﬁcation

model. Tables 2, 3, 4, and 5 display the confusion

matrix for each model. The obtained results enforce

the same conclusions presented in the previous anal-

ysis.

Table 2: Confusion Matrix - SVM.

Fresh Rotten

Fresh 96.9 % 3.1 %

Rotten 18.1 % 81.9 %

In a further evaluation, we also analyzed the classiﬁ-

IDiSSC: Edge-computing-based Intelligent Diagnosis Support System for Citrus Inspection

689

Table 3: Confusion Matrix - KNN.

Fresh Rotten

Fresh 94.6 % 5.4 %

Rotten 4.5 % 95.5 %

Table 4: Confusion Matrix - Random Forest.

Fresh Rotten

Fresh 95.1 % 4.9 %

Rotten 2.2 % 97.8 %

Table 5: Confusion Matrix - MLP (ANN).

Fresh Rotten

Fresh 99.0 % 1.0 %

Rotten 1.0 % 99.0 %

cation using multiple classes. As presented, our com-

posed dataset has images of fresh oranges, oranges

with canker, Greening, and black spot. Among the

tested methods, the artiﬁcial neural network also pre-

sented the best results for the oranges diagnosis. Ta-

ble 6 displays the obtained results on this stage of the

research.

Table 6: Models Test - Multiple Classes.

Precision Recall F1-Score Support

SVM Results:

Fresh 1.00 1.00 1.00 149

Canker 0.91 0.88 0.89 98

Greening 1.00 1.00 1.00 179

Black Spot 0.89 0.92 0.91 110

Accuracy = 96.08% 536

KNN Results:

Fresh 1.00 0.98 0.99 141

Canker 0.88 0.92 0.90 92

Greening 1.00 1.00 1.00 208

Black Spot 0.90 0.88 0.89 95

Accuracy = 96.0% 536

Random Forest Results:

Fresh 1.00 1.00 1.00 127

Canker 0.94 0.87 0.90 105

Greening 1.00 1.00 1.00 191

Black Spot 0.88 0.95 0.91 113

Accuracy = 96.26% 536

MLP(ANN) Results:

Fresh 1.00 1.00 1.00 154

Canker 0.94 0.96 0.95 96

Greening 1.00 1.00 1.00 188

Black Spot 0.96 0.94 0.95 98

Accuracy = 98.13% 536

To conﬁrm the obtained results, we also evaluated the

confusion matrix for each classiﬁcation model. Ta-

bles 7, 8, 9, and 10 display the confusion matrix for

each model. Again, the obtained results enforce the

same conclusions presented in the previous analysis.

After analyzing the software performance, we also

need to evaluate the hardware aspects.

Table 7: Confusion Matrix - SVM.

Fresh Canker Greening Black Spot

Fresh 97.0 % 1.0 % 1.0 % 1.0 %

Canker 1.0 % 88.6 % 1.0 % 10.4 %

Greening 1.0 % 1.0 % 98.3 % 1.0 %

Black Spot 1.0 % 9.27 % 1.0 % 87.8 %

Table 8: Confusion Matrix - KNN.

Fresh Canker Greening Black Spot

Fresh 97.8 % 1.0 % 1.0 % 2.1 %

Canker 1.0 % 86.7 % 1.0 % 7.4 %

Greening 1.0 % 1.0 % 98.5 % 1.0 %

Black Spot 1.0 % 11.2 % 1.0 % 89.4 %

Table 9: Confusion Matrix - Random Forest.

Fresh Canker Greening Black Spot

Fresh 97.7 % 1.0 % 1.0 % 1.0 %

Canker 1.0 % 88.7 % 1.0 % 11.38 %

Greening 1.0 % 1.0 % 98.5 % 1.0 %

Black Spot 1.0 % 5.7 % 1.0 % 86.9 %

Table 10: Confusion Matrix - MLP (ANN).

Fresh Canker Greening Black Spot

Fresh 98.0 % 1.0 % 1.0 % 1.0 %

Canker 1.0 % 92.0 % 1.0 % 4.0 %

Greening 1.0 % 1.0 % 98.4 % 1.0 %

Black Spot 1.0 % 6.0 % 1.0 % 93.8 %

5.2 Hardware Performance Test

After testing each algorithm’s performance regarding

the prediction processes, we also needed to test each

method’s performance, considering the hardware con-

straints. As presented before, in the algorithm pro-

cess, there are three main stages:

1. Acquire image;

2. Extract feature vector;

3. Predict class;

To understand the impact of each stage, we evalu-

ated the average behavior concerning three different

hardware conﬁgurations: a high-performance desk-

top, identiﬁed as DT, an average-performance per-

sonal laptop computer, identiﬁed as NB, and an edge

computing device, identiﬁed as PI3. We presented the

conﬁguration of each element in Section 4.

• The results obtained from the SVM display that:

– The image acquisition stage took 2.0 ± 1.0 ms

in the DT machine, 2.5 ± 1.2 ms in NB ma-

chine, and 12.6 ± 5.8 ms in PI3 device;

– The feature extraction process took 0.4 ± 0.1

ms in DT, 0.5 ± 0.2 ms in NB, and 4.2 ± 1.2

ms in PI3.

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

690

– Finally, the classiﬁcation stage took 13.4 ± 1.7

ms in DT, 17.3 ± 3.8 ms in NB, and 195.4 ±

11.3 in PI3;

• The results obtained from the KNN display that:

– The image acquisition stage took 2.0 ± 0.9 ms

in the DT machine, 2.5 ± 1.1 ms in NB ma-

chine, and 12.7 ± 5.7 ms in PI3 device;

– The feature extraction process took 0.4 ± 0.1

ms in DT, 0.5 ± 0.2 ms in NB, and 4.4 ± 1.4

ms in PI3.

– Finally, the classiﬁcation stage took 14.6 ± 5.8

ms in DT, 19.1 ± 10.8 ms in NB, and 553.0 ±

88.6 in PI3;

• The results obtained from the Random Forest

display that:

– The image acquisition stage took 1.9 ± 0.9 ms

in the DT machine, 2.5 ± 1.1 ms in NB ma-

chine, and 12.7 ± 5.8 ms in PI3 device;

– The feature extraction process took 0.4 ± 0.1

ms in DT, 0.6 ± 0.3 ms in NB, and 4.0 ± 1.0

ms in PI3.

– Finally, the classiﬁcation stage took 10.3 ± 1.6

ms in DT, 13.6 ± 3.6 ms in NB, and 123.1 ±

13.5 in PI3;

• The results obtained from the MLP (ANN) dis-

play that:

– The image acquisition stage took 2.0 ± 0.9 ms

in the DT machine, 11.5 ± 28.7 ms in NB ma-

chine, and 22.7 ± 88.8 ms in PI3 device;

– The feature extraction process took 0.4 ± 0.1

ms in DT, 1.2 ± 8.0 ms in NB, and 4.6 ± 2.5

ms in PI3.

– Finally, the classiﬁcation stage took 11.0 ± 1.8

ms in DT, 18.9 ± 13.0 ms in NB, and 135.1 ±

15.1 in PI3;

6 DISCUSSION

In this work, we present a system for classifying or-

anges using an embedded application. For this mat-

ter, we proposed techniques using machine learning

functions to determine whether the fruits show signs

of infection by diseases. We also examined the main

constraints regarding the employed hardware, given

an embedded system application’s perspective.

Our system aims to aid technicians and small crop

processing plants in diagnosing diseases in the orange

crop. For this matter, we propose the usage of an em-

bedded system incorporated with a machine learning

algorithm. This appliance uses a “pseudospectrum”

extracted from the HSV color space.

To test this proposal, at ﬁrst, we created multiple

algorithms with a labeled dataset containing images

from fresh and rotten oranges (Kalluri, 2018). We

considered SVM, KNN, Random Forest, and MLP as

possible candidates to integrate the solution. To test

the system feasibility, we tested the algorithms using

machine learning metrics. We also evaluated the tim-

ing constraints for checking the efﬁciency using vari-

ous hardware conﬁgurations, including an edge com-

puting device.

All proposed methods had over 92% accuracy

when separating the data. The Random Forest and

MLP algorithms had the best and most balanced mod-

els. The results for the classiﬁcation using multiple

classes enforce the feasibility of this system using the

proposed algorithms. Also, from the hardware eval-

uation, we veriﬁed that the edge device could per-

form approximately 4.95 predictions/s with the SVM,

1.78 predictions/s using KNN, 7.63 predictions/s us-

ing the Random Forest, and 6.17 predictions/s using

the MLP.

These results enforce that the leading candidate

models for integrating the proposed solution are the

MLP and the Random Forest, given their performance

in the prediction process and predicting using the em-

bedded hardware. Future work in this context must

consider testing the system in actual ﬁeld applica-

tions, integrating it into the citrus fruits’ productive

process. In this perspective, future research should

consider measuring aspects related to embedded sys-

tems, such as energy consumption and hardware re-

source proﬁling.

ACKNOWLEDGEMENTS

The authors would like to thank CAPES, CNPq,

Instituto Federal de Educac¸

ao, Ci

encia e Tecnolo-

gia de Minas Gerais and the Universidade Fed-

eral de Ouro Preto for supporting this work. This

study was ﬁnanced in part by the Coordenac¸

ao de

Aperfeic¸oamento de Pessoal de N

ıvel Superior -

Brasil (CAPES) - Finance Code 001.

REFERENCES

4vision (2019). Como a vis

ao computacional est

a revolu-

cionando a gest

ao de qualidade na ind

ustria.

Bhavani, R. and Wiselin Jiji, G. (2018). Image registration

for varicose ulcer classiﬁcation using knn classiﬁer.

IDiSSC: Edge-computing-based Intelligent Diagnosis Support System for Citrus Inspection

691

International Journal of Computers and Applications,

40(2):88–97.

Blasco, J., Cubero, S., and Molt

o, E. (2016). Computer

Vision Technology for Food Quality Evaluation. San

Diego: Academic Press, 2 edition.

C¸ alik, R. C. and Demirci, M. F. (2018). Cifar-10 image

classiﬁcation with convolutional neural networks for

embedded systems. In 2018 IEEE/ACS 15th Interna-

tional Conference on Computer Systems and Applica-

tions (AICCSA), pages 1–2. IEEE.

CitrusBr (2020). Laranja e suco a fruta.

da Rosa, A. L. (2019). Classiﬁcac

ao de imagens de frutas

utilizando aprendizado de m

aquina.

Ebrahimi, M., Khoshtaghaza, M., Minaei, S., and Jamshidi,

B. (2017). Vision-based pest detection based on svm

classiﬁcation method. Computers and Electronics in

Agriculture, 137:52–58.

Fundecitrus (2019). Sete erros no controle da pinta preta.

Gottwald, T. R., Graham, J. H., and Schubert, T. S. (2002).

Citrus canker: the pathogen and its impact. Plant

Health Progress, 3(1):15.

Guazzelli, A., Zeller, M., Lin, W.-C., Williams, G., et al.

(2009). Pmml: An open standard for sharing models.

The R Journal, 1(1):60–65.

Hashemi, S. M. R., Hassanpour, H., Kozegar, E., and Tan,

T. (2020). Cystoscopic image classiﬁcation based on

combining mlp and ga. International Journal of Non-

linear Analysis and Applications, 11(1):93–105.

Huang, K., Li, S., Kang, X., and Fang, L. (2016). Spectral–

spatial hyperspectral image classiﬁcation based on

knn. Sensing and Imaging, 17(1):1.

JAI (2020). Prism-based multispectral cameras empower

high speed fruit sorting.

Jaskolka, K., Seiler, J., Beyer, F., and Kaup, A. (2019). A

python-based laboratory course for image and video

signal processing on embedded systems. Heliyon.

Kalluri, S. R. (2018). Fruits fresh and rotten for classiﬁca-

tion apples oranges bananas.

Kumar, M. (2017). Implementing a binary classiﬁer in

python.

LeCun, Y. (2019). 1.1 deep learning hardware: Past,

present, and future. In 2019 IEEE International

Solid-State Circuits Conference-(ISSCC), pages 12–

19. IEEE.

Mathew, A. R. and Anto, P. B. (2017). Tumor detection

and classiﬁcation of mri brain image using wavelet

transform and svm. In 2017 International Conference

on Signal Processing and Communication (ICSPC),

pages 75–78. IEEE.

Neves, M. F. and Trombin, V. G. (2017). Anu

ario da Citri-

cultutura 2017. citrusbr, S

ao Paulo, 1 edition.

Padol, P. B. and Yadav, A. A. (2016). Svm classiﬁer based

grape leaf disease detection. pages 175–179.

Pipatnoraseth, T., Phognsuphap, S., Wiratkapun, C., Tana-

wongsuwan, R., Sajjacholapunt, P., and Shimizu, I.

(2019). Breast microcalciﬁcation visualization using

pseudo-color image processing. In 2019 12th Biomed-

ical Engineering International Conference (BME-

iCON), pages 1–5. IEEE.

Puchkov, E. et al. (2016). Image analysis in microbiology:

a review. Journal of Computer and Communications,

4(15):8.

Puchkov, E. and McCarren, M. (2011). Assessment of

the distribution of nucleic acid intercalators in yeast

cells by pseudospectral image analysis. Biophysics,

56(4):651.

Putra, K. T., Hariadi, T. K., Riyadi, S., and Chamim, A.

N. N. (2018). Feature extraction for quality modeling

of malang oranges on an automatic fruit sorting sys-

tem. In 2018 2nd International Conference on Imag-

ing, Signal Processing and Communication (ICISPC),

pages 74–78. IEEE.

Rauf, Tayyab, H., Saleem, B. A., Lali, M. I. U., khan, a.,

Sharif, M., and Bukhari, S. A. C. (2009). A citrus

fruits and leaves dataset for detection and classiﬁca-

tion of citrus diseases through machine learning. 2.

Rong, D., Ying, Y., and Rao, X. (2017). Embedded vision

detection of defective orange by fast adaptive light-

ness correction algorithm. Computers and Electronics

in Agriculture, pages 48–59.

Rotondo, T., Farinella, G. M., Chillemi, A., Ferlito, F., and

Battiato, S. (2018). A digital countryside notebook for

smart agriculture and oranges classiﬁcation. In ICETE

(1), pages 547–551.

Shankar, K., Zhang, Y., Liu, Y., Wu, L., and Chen, C.-H.

(2020). Hyperparameter tuning deep learning for di-

abetic retinopathy fundus image classiﬁcation. IEEE

Access.

Silva, C. F. and Siebra, C. A. (2017). An investigation on the

use of convolutional neural network for image classi-

ﬁcation in embedded systems. In 2017 IEEE Latin

American Conference on Computational Intelligence

(LA-CCI), pages 1–6. IEEE.

Soini, C. T., Fellah, S., and Abid, M. R. (2019). Green-

ing infection detection (cigid) by computer vision and

deep learning.

Tu, B., Wang, J., Kang, X., Zhang, G., Ou, X., and Guo,

L. (2018). Knn-based representation of superpixels

for hyperspectral image classiﬁcation. IEEE Journal

of Selected Topics in Applied Earth Observations and

Remote Sensing, 11(11):4032–4047.

USDA APHIS (2018). Citrus greening.

Valarmathie1, P., Sivakrithika, V., and Dinakaran, K.

(2016). Classiﬁcation of mammogram masses using

selected texture, shape and margin features with mul-

tilayer perceptron classiﬁer.

Wasule, V. and Sonar, P. (2017). Classiﬁcation of brain mri

using svm and knn classiﬁer. In 2017 Third Interna-

tional Conference on Sensing, Signal Processing and

Security (ICSSS), pages 218–223. IEEE.

Yin, S., Bi, X., Niu, Y., Gu, X., and Xiao, Y. (2017). Hyper-

spectral classiﬁcation for identifying decayed oranges

infected by fungi. Emirates Journal of Food and Agri-

culture, pages 601–609.

ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

692