Jour nal o f P la n t Pro t e ct io n R e s e a rc h eISSN 1899-007X

ORIGINAL ARTICLE

Tomato fruit disease detection based

on improved single shot detection algorithm
Benedicta Nana Esi Nyarko* , Wu Bin, Zhou Jinzhi , Justice Odoom ..
School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan, China

Vol. 63, No. 4: 405–417, 2023 Abstract

DOI: 10.24425/jppr.2023.146877 The tomato crop is more susceptible to disease than any other vegetable, and it can be
infected with over 200 diseases caused by different pathogens worldwide. Tomato plant
Received: May 17, 2023 diseases have become a challenge to food security globally. Currently, diagnosing and pre-
Accepted: July 06, 2023 venting tomato plant diseases is a challenge due to the lack of essential methods or tools.
Online publication: October 02, 2023 The traditional techniques of detecting plant disease are arduous and error-prone. Utiliz-
ing precise or automatic detection methods in spotting early plant disease can improve
*Corresponding address: the quality of food production and reduce adverse effects. Deep learning has significantly
[email protected] increased the recognition accuracy of image classification and object detection systems in
Responsible Editor: recent years. In this study, a 15-layer convolutional neural network is proposed as the back-
Rafał Kukawka bone for single shot detector (SSD) to improve the detection of healthy, and three classes
of tomato fruit diseases. The proposed model performance is compared with ResNet-50,
AlexNet, VGG 16, and VGG19 as the backbone for Single shot detector. The findings of
the experiment showed that the proposed CNN-SDD achieved 98.87% higher detection
accuracy, which outperformed state-of-the-art models.
Keywords: convolutional neural network, deep learning, feature extraction, model back-
bone, plant disease detection, single shot detector algorithm

Introduction

Modern technologies can produce enough food to health are all harmed by the extensive use of these
meet the requirements of over 7 billion people (He- chemical techniques. Additionally, these techniques
mathilake and Gunathilake 2022). Yet, several issues, raise production costs. Plant disease management is
such as decreasing pollinators, climate change, and effective when diseases are detected at an early stage.
plant disease continually pose a threat to food stability Lycopersicon esculentum (Mohan et al. 2023) (To-
(Hofman-Bergholm 2023). In addition to endangering mato) previously referred to as Solanum Lycopersi-
the safety of the world’s food supply, plant diseases can con L (Leite and Fialho 2018) belongs to the family
have tragic effects on smallholder farmers whose liveli- Solanaceae (Knapp and Peralta 2016) and emanates
hoods depend on robust crops. Several initiatives have from the Andes region of South America. Cur-
been created to stop crop loss from diseases (Khan et rent global research indicates that approximately
al. 2023). The integrated pest management approach 182 million tons of tomato crops are cultivated from
has largely substituted conventional techniques for ap- 5 million hectares of land (Caruso 2022, Bhujel et al.
plying insecticides in the past 10 years (Golan et al. 2022). With 0.2 million hectares, it is the lead-
2023). The environment, human health, and plant ing vegetable with the largest yield in southern
406 Journal of Plant Protection Research 63 (4), 2023

Europe. Turkey, Egypt, and Italy are the leading pro- very innovative methods to aid in the identification of
ducers. This is a blooming species, most frequently diseases because of their computing capacity. For im-
a sprawling and nightshade plant that is usually grown age-based research, CNN models are frequently used.
for its edible fruit (Liu et al. 2022). They are effective at extracting high-level features from
Phonological changes in the tomato plant can in- images (Ma et al. 2023). This study focused on detect-
clude aberrant growth, pigmentation, spots, deformi- ing tomato fruit disease by proposing a 15-layer CNN
ties, wilting, desiccation, and necrosis on the leaves architecture as a base network for single shot detector
(Humbal and Pathak 2023, Kremneva et al. 2023; (SSD) to extract high features from tomato fruit. The
Nkongho et al. 2023; Sreedevi and Manike 2023). To- article is organized as follows: the related work of this
mato disease control is best when all available methods research, followed by the materials and methods used
are used (Albattah et al. 2022). Cultural operations are for this study, the proposed backbone, and finally, the
aimed at preventing disease or impeding its occur- experimental setup, results and discussion.
rence. Common tomato diseases include: black mold,
rot, crown and root rot, spotted wilt virus, radial rings
cankers, mildews, blights, and many more (Ates et al. Materials and Methods
2019, Gatahi 2020). Black mold fungus, Alternaria al-
ternate, is one of the most common fungi experienced.
It breeds dead organic matter whenever moisture The proposed model architecture, hardware, data-
is present and can be found on fruit and fermenting set, and software resources used for this research are
leaves in tomato fields before the fruit ripens (Rashid briefly discussed in this section. A single short shot
and Shoala 2020). Alternaria fungi is also responsi- detector with a proposed CNN base network was used
ble for early blight, and stem canker, which are often in this study for tomato fruit disease detection. SSD
confused with black mold (Sánchez et al. 2022). These was proposed to accurately locate the infected area of
two diseases are caused by A. solani and A. alternata the tomato fruit. The experiments were carried out on
f.s. It is an economically destructive disease on tomato a Lenovo laptop with an Intel Core i7 2.50 GHz proces-
plants (Nazari et al. 2022). Controlling the most prev- sor and NVidia GeForce GTX 860M GPU.
alent type of Fusarium crown and root-rot disease is
very challenging because it is a soil-borne disease that
Data collection and pre-processing
economically restricts the production of greenhouse
tomatoes. Due to the spread of microconidia, espe- Tomato fruit images were collected from the internet
cially in greenhouses, FORL causes recurrent infec- and screened carefully to correspond with the disease
tions throughout the growing season and can cause up classes. This included black mold diseases, radial ring
to 90% crop losses in greenhouse tomato cultivation. diseases, spotted wilt diseases, and healthy tomato
(Özbay et al. 2002). Commercially viable, agent-resist- fruits. A training set and a testing set were created
ant indigenous varieties with adequate resistance to from the dataset in the following proportions: 8 : 2.
Fusarium crown and root-rot induced by FORL have Figure 1 shows samples of the tomato fruit classes in
not yet been developed (Ozbay and Newman 2004). the dataset.
A Gram-positive actinobacterium named Clavibac- A total of 2500 images were obtained and a total of
ter michiganensis causes bacterial canker of tomatoes, 650 images were obtained for the radial ring class rep-
a disease affecting millions of people worldwide (Per- resenting 26% of the total data. Seven hundred images
itore-Galve et al. 2020). Due to its rapid migration were obtained for the healthy class representing 28%
through the plant’s vascular system and its ability to of the total data, 600 images for the spotted wilt class
induce systemic symptoms, bactericidal agents are in- representing 24% of the total data, and 550 images
sufficient to manage this disease for the black mold class representing 22% of the total
Image processing and computer vision is key for class. The data distribution is graphically presented in
detecting plant diseases (Ouhami et al. 2021). It is Figure 2.
a technique that gathers plant images using comput- To make the tomato fruit detectable by SSD, the
er vision tools and uses those images to determine MATLAB Image labeler app was used to annotate the
whether or not pests and diseases are present (Vishnoi datasets. The app offers a simple method for creating
et al. 2021). Deep learning has recently been used in interactively a variety of shapes to classify as regions
the evaluation and identification of plant diseases, fa- of interest (ROI). A rectangle shape was used to mark
cilitating early disease detection and diagnosis, there- the disease spot’s size and shape on the tomato fruit.
by quickening the development of novel plant disease Figure 3 shows sample annotated data in the MATLAB
technology (Thakur et al. 2023). Deep learning offers image labeler.
 Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 407

A B

C D

Fig. 1. Sample tomato images: A – black mold disease; B – spotted wilt virus; C – radial ring disease; D – Healthy tomato fruit

The input layer, convolution layer, pooling layer, full

connection layer, and output layer make up the con-
volutional neural network design (Ji et al. 2022). One
model repeatedly alternates between the convolution
layer and the pooling layer, and when the neurons of
the convolution layer are connected to the neurons of
the pooling layer, full links are not necessary. In image
identification, CNN is frequently used because of its
magnificent potential (Gaba et al. 2022). CNNs have
been used in the framework to manage the extracted
image data that was gathered during the feature ex-
traction stage (Ghazal 2022). The distinctive features
of the affected tomato fruit are accurately extracted by
the feature extraction layer of the proposed neural net-
works to achieve high recognition rates for tomato dis-
eases (Bouni et al. 2023; Thakur et al. 2023). The per-
formance of the proposed model was compared with
Fig. 2. Graphical presentation of dataset
the existing state-of-art models such as ResNet-50,
AlexNet, and VGG19. Equation 1 illustrates a CNN
layer-by-layer pass operation:

→ w1 → x2 → ... → xn–1 → wn–1 → xn → wn → z, (1)

Convolutional neural network
CNNs (Haar et al. 2023), or convolutional neural net- where: x1 – represents the input image, w1 – denotes
works, have a complicated network topology and are the parameter in the first layer, x2– denotes the output
capable of performing convolution operations. There of the first layer, and xn – denotes the output of the
are two broad categories of CNN-based image detec- final CNN layer. A loss layer is the final layer. The dif-
tion systems. In the first category, samples consisting ference between the CNN prediction xn and the target
of several candidate frames are created and then classi- can be measured using a loss function, assuming that
fied using CNN models. Several domains have success- y is the corresponding target value for the input x1. For
fully used deep neural networks. For the best possible instance, Equation 2 represents a straightforward loss
image detection, CNN was used (Nyarko et al. 2022). function that could be:
 408 Journal of Plant Protection Research 63 (4), 2023

A B

C D

Fig. 3. Samples of annotated data in MATLAB image labeler. (A) Sample annotated healthy tomato in MATLAB image labeler.
(B) Sample Annotated Spotted Wilt infected tomato in MATLAB image labeler. (C) Sample annotated radial ring affected
Tomato in MATLAB image labeler. (D) A sample annotated black mold infected tomato in MATLAB image labeler.

z = 1/2║y – xn║2 , (2) Single shot detection algorithm (SSD)

2-5
‌‌‌where: y – denote the ground truth value for the input, In a single pass, the SSD model identifies objects with 2-5 2-5
and xn – denote the difference between 2-5 the CNN pre- incredible detection accuracy, which saves a significant
2-5
1
amount
𝑧𝑧𝑧𝑧 = 2 ‖𝑦𝑦𝑦𝑦of − time 𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛 ‖2 , (Li et al. 2023, 2 Vig et al. 2023). It yields
diction. The chain rule and vector calculus are essen-
tial to the CNN learning 1process. If If 𝑦𝑦𝑦𝑦y ∈ ℝ accurate predictions at diverse levels from feature maps
𝑛𝑛𝑛𝑛 ‖2 𝑧𝑧𝑧𝑧 =is a‖scalar
ℝ B
B 1
𝑧𝑧𝑧𝑧 = ‖ 𝑦𝑦𝑦𝑦 − 𝑥𝑥𝑥𝑥 , 𝑦𝑦𝑦𝑦 −2𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛 ‖2 , 2
the partial derivative of 1 a to y is a vector if a is a func-
2 and achieves 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 high detection accuracy by directly divid-
𝑧𝑧𝑧𝑧 B= ‖𝑦𝑦𝑦𝑦2− 𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛 ‖2 , �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �= 3
2-52-5
2 ing 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
predictions by aspect ratio (Guravaiah et al. 2023).
2
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
B tion of y,Ifand 𝑦𝑦𝑦𝑦 ∈ it isℝdefined in Equation 3 as: 2
If 𝑦𝑦𝑦𝑦 ∈ ℝ B 1 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 input images with low resolutions,
𝑧𝑧𝑧𝑧 = 2 ‖𝑦𝑦𝑦𝑦 − 𝑥𝑥𝑥𝑥 𝑛𝑛𝑛𝑛 ‖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 2
2Even of on these tech-
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
If 𝑦𝑦𝑦𝑦 ∈ ℝ , The𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖 𝑡𝑡𝑡𝑡ℎ element
𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦
, a vector of the same size as y, is . Given that
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 , � � = niques produce high accuracy and simple end-to-end 3
��𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕(3) 𝑖𝑖𝑖𝑖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 3
=� =
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
If 𝑦𝑦𝑦𝑦 ∈ ℝB �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 3
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 � = �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕� and training.
y is a function TheofSSD x given works
that 𝑥𝑥𝑥𝑥by ∈ employing
ℝ𝑊𝑊𝑊𝑊 is anotherconvolutional
vector
3 𝑡𝑡𝑡𝑡ℎ 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 2-5�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇of� the networks to produce a large 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕number of bounding box-
The 𝑖𝑖𝑖𝑖1 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
1𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 element
𝑛𝑛𝑛𝑛 ‖2 𝑛𝑛𝑛𝑛 of2 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦, a vector 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 same
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 size as y, is .4Given that
The The𝑖𝑖𝑖𝑖 𝑡𝑡𝑡𝑡ℎ 𝑡𝑡𝑡𝑡ℎelementof
𝑖𝑖𝑖𝑖element element 𝑧𝑧𝑧𝑧of=𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 of a‖=
,2�𝑧𝑧𝑧𝑧𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑦𝑦𝑦𝑦�, a=
vector −vector 𝑥𝑥𝑥𝑥 of the
𝑦𝑦𝑦𝑦𝑖𝑖𝑖𝑖 − 𝑥𝑥𝑥𝑥 ‖ , ,
of same the same
size as size y, 2 as 3 𝑖𝑖𝑖𝑖y,
= 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝑖𝑖𝑖𝑖ofis
.
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
. Given that
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 2
The vector of the same size as y, is 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 . Given that 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 2 ‖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 es various
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
fixed sizes and assess whether or not an
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑇𝑇𝑇𝑇 instance of an object 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
. class is present in each box. Af-
me
∈ sizeB as y, is = .�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
is Given
Given 𝑇𝑇𝑇𝑇 that
𝑇𝑇𝑇𝑇 that
�is𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 � =𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �x and and yyis1is a afunction −function , is of xthe given that
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑥𝑥𝑥𝑥 ∈ ℝ
𝑊𝑊𝑊𝑊
is another vector their job,
If 𝑦𝑦𝑦𝑦ℝ
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
∈ ℝ��B𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑥𝑥𝑥𝑥∈𝑛𝑛𝑛𝑛 ‖ℝ 2 𝑊𝑊𝑊𝑊
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑇𝑇𝑇𝑇 �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑡𝑡𝑡𝑡ℎ and y a function of
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 given 𝑧𝑧𝑧𝑧 = that ‖𝑦𝑦𝑦𝑦 𝑥𝑥𝑥𝑥 ter
another 2 convolutional
vector networks have finished
� = 𝑖𝑖𝑖𝑖The
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖� element
� and y is of a function , a vector of x of given
2
the same
that 𝑥𝑥𝑥𝑥 size ∈
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 ℝ as 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 y,
𝑊𝑊𝑊𝑊 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕is
is another . Given vector that
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦 =
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 a non-maximum
. 5
of x given that If 𝑦𝑦𝑦𝑦x𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 ∈ℝ ℝBW𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 is 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 another vector. 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 The 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 partial
𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 suppression step then produces the
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇�
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 =� = 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕. 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 � = . 4 3 4 al. 2022).
that 𝑥𝑥𝑥𝑥 ∈ ℝ derivative 𝑊𝑊𝑊𝑊
is another
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 of𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 y �vector
concerning
𝑇𝑇𝑇𝑇𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 �=𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 � 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 �x is =
𝑖𝑖𝑖𝑖 .then expressed � 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕in
� = Equa-
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 final
4 3 3 𝑊𝑊𝑊𝑊
detections (Shi et
= �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑇𝑇𝑇𝑇and 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕of x𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕given
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑇𝑇𝑇𝑇4�as:
tion 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 y is a function
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝑖𝑖𝑖𝑖 that 𝑥𝑥𝑥𝑥 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 ∈The ℝ head is another and the backbone vector make up the single shot
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 The . 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 detection .
sizeisas 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 y,design. Thethat backbone layer is utilized by the
, a𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑡𝑡𝑡𝑡ℎ
The element of𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 of vector of the same size as y, is. Given that
𝑡𝑡𝑡𝑡ℎ
element 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 of𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 , a vector of the same .𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
Given
4 𝑖𝑖𝑖𝑖The 𝑖𝑖𝑖𝑖 𝑡𝑡𝑡𝑡ℎ 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕element
𝑖𝑖𝑖𝑖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
, a vector 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦of . the same size
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
as y, is
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
. Given that
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦� = . feature 4 map 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 generator (Wei et al. 2023), a standard im-
𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 . 𝜕𝜕𝜕𝜕𝑦𝑦𝑦𝑦𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 (4)
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 5 𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
= 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 = . 5
𝑇𝑇𝑇𝑇
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑇𝑇𝑇𝑇 �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇
� = � 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
� and
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 y is a function
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇
of x given age classification network that has been trained. The
that 𝑥𝑥𝑥𝑥 ∈ ℝ 𝑊𝑊𝑊𝑊
is another vector
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 = 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 .
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 5ℝ𝑊𝑊𝑊𝑊
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑇𝑇𝑇𝑇 ��= ��𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 The
= � � and
H ×𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 D
� 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 yand
matrix 𝑇𝑇𝑇𝑇 is 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕ay𝑇𝑇𝑇𝑇function
that
is makes
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕a𝑇𝑇𝑇𝑇 function �of
up �x =
this
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 of
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 given partial
x
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
. .given that 𝑥𝑥𝑥𝑥that
deri­ ∈final 𝑥𝑥𝑥𝑥removed, ∈ is4ℝanother
image classification
𝑊𝑊𝑊𝑊
is another vector layer produced by the model
vector
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 is leaving the extracted feature maps. The
va­tive has an entry at the point where the ith row and 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖

5 the ith column . SSD head is constructed by stacking convolutional lay-

𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
�𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 � �𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑇𝑇𝑇𝑇meet =𝑇𝑇𝑇𝑇is= .
𝑇𝑇𝑇𝑇 In
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 ..
a𝑇𝑇𝑇𝑇chain-like argument, 4 5
� ers and is4placed on top of the backbone model (Liu
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝑖𝑖𝑖𝑖𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 =
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 = 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 . 5
it is simple to see that a is a function 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇of x: 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇one𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 function et al. 2023a). The SSD algorithm divides each input im-
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖
maps x to y, and another function maps y to. a. One can age into grids of different sizes, and at each grid, detec-
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖.
compute using the chain rule in Equation 5: tion for various classes and aspect ratios is performed.
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 A score is assigned to each of these grids to represent
=𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇= 𝑇𝑇𝑇𝑇. 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
. 5
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑇𝑇𝑇𝑇 (5) how well 5an object works in that particular grid. The
final detection is then extracted using non-maximum
 Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 409

suppression from the collection of overlapping detec- Input layer

tions (Liu et al. 2023b). The SSD model’s architecture is It acts as an input element for the neural network. Each
presented in Figure 4. input layer feature passes its assigned value to each
neuron in the first hidden layer in a top-to-bottom se-
quence. The hidden layer neuron then adds each value
Proposed backbone (CNN)
to the weight vector that corresponds to it, sums the
The most significant area of research in the field of deep multiplied values, applies its activation function to this
learning has always been the design of a high-quality, total, and passes the value calculated by the activation
high-efficiency expressive network architecture. The function to the following layer. Input color channels
majority of current network design approaches con- are used to encrypt images. Each color level in the
centrate on how to combine features extracted from color channel at a specific location represents image
various layers and how to create computing units that data. RGB, which means red, blue, and green, is the
can extract these features efficiently, increasing the ex- most popular. The information contained in the image
pressiveness of the network. A key design section of is the intensity of each channel color across the width
a one-stage detector model is the backbone, which de- and height of the input image. The blend of these three
termines the quality of image feature extraction. It also colors forms a color pattern.
affects the seceding object detection, recognition, and
object classification. Figure 5 shows the proposed CNN Convolutional layer
architecture and Figure 6 shows the structure of the pro- The convolutional layer examines each neural network
posed backbone with a single shot detector algorithm. patch to derive more abstract properties. It performs
The basic components of the proposed CNN used convolution on the input using kernels, which were fil-
as the backbone for SSD in detecting tomato fruit dis- ters in the past. Convolution appears to be dot products
eases are described as follows between the filtering and the region they pass through

Fig. 4. Structure of single shot detector (SSD) architecture

Fig. 5. Proposed CNN architecture

 410 Journal of Plant Protection Research 63 (4), 2023 6-8
6-8

𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙

𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1,𝑑𝑑𝑑𝑑 = ∑𝐻𝐻𝐻𝐻 𝐵𝐵𝐵𝐵
𝑖𝑖𝑖𝑖=0 ∑𝑖𝑖𝑖𝑖=0 ∑𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙, X 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1, 𝑑𝑑𝑑𝑑 ,
𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖
𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙
𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1,𝑑𝑑𝑑𝑑 = ∑𝐻𝐻𝐻𝐻 𝐵𝐵𝐵𝐵
𝑖𝑖𝑖𝑖=0 ∑𝑖𝑖𝑖𝑖=0 ∑𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙, X 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1, 𝑑𝑑𝑑𝑑 ,
𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖
𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1
0 ≤ d, < D = 𝐷𝐷𝐷𝐷 , for 𝑖𝑖𝑖𝑖 , 𝑗𝑗𝑗𝑗 satisfying 0 ≤ 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 < 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 −
0𝐵𝐵𝐵𝐵 d, < D
≤𝑙𝑙𝑙𝑙+1 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1
in=this , for 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 , 𝑗𝑗𝑗𝑗𝑥𝑥𝑥𝑥𝑙𝑙𝑙𝑙+1
equation, 𝑙𝑙𝑙𝑙 satisfying 0 ≤ 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 < 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 − 𝐻𝐻𝐻𝐻 +𝑙𝑙𝑙𝑙 1, <
refers to the element of 𝑥𝑥𝑥𝑥 ind
𝑙𝑙𝑙𝑙
𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖+1,𝑑𝑑𝑑𝑑
𝐵𝐵𝐵𝐵 𝑙𝑙𝑙𝑙+1
in this equation, 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙𝑖𝑖𝑖𝑖+1,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 refers to the element of 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 indexed 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙+1
𝑖𝑖𝑖𝑖

Fig. 6. Structure of proposed backbone with single shot detector (SSD)

6-8= max�0, 𝑥𝑥𝑥𝑥

𝑙𝑙𝑙𝑙 �,
𝑦𝑦𝑦𝑦
𝑦𝑦𝑦𝑦𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑
= max�0, 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 𝑖𝑖𝑖𝑖.𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 �, 𝑖𝑖𝑖𝑖.𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑
in reality. To reduce the variable value and extract the ReLu << layers do𝑙𝑙𝑙𝑙+1
not0carry any𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1
parameters, so this
6-8
𝑙𝑙𝑙𝑙
0layer <0 idoes i𝐻𝐻𝐻𝐻<
𝑙𝑙𝑙𝑙+1 𝐻𝐻𝐻𝐻
, 0 < 𝐽𝐽𝐽𝐽,parameter
<<𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙𝐽𝐽𝐽𝐽 =<𝐵𝐵𝐵𝐵learning. = 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1
, 𝜕𝜕𝜕𝜕𝑎𝑎𝑎𝑎𝑑𝑑𝑑𝑑 0 ,≤𝜕𝜕𝜕𝜕𝑎𝑎𝑎𝑎𝑑𝑑𝑑𝑑
≤0𝐷𝐷𝐷𝐷≤
𝑑𝑑𝑑𝑑Equa- 𝑙𝑙𝑙𝑙 𝑑𝑑𝑑𝑑 ≤
𝑙𝑙𝑙𝑙
≤ 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1𝐷𝐷𝐷𝐷. ≤ 𝐷𝐷𝐷𝐷
key traits, a convolution layer was used. Rotation in- not require From
terpretability, scale interpretability, and interpretation tion 7 we obtain Equation 8 as: 9-11
interpretability were all included in the convolution 𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙
layer. Both the generalization concept and the over-fit- = �𝑥𝑥𝑥𝑥 = �𝑥𝑥𝑥𝑥
𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖𝑙𝑙𝑙𝑙,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖𝑙𝑙𝑙𝑙,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 > 0�,
> 0�, (8) 8
ting issue were added to the fundamental framework.
The operation of the convolution layer in CNN can be where ‖. ‖ –‖.the
where
where ‖ – the
the indicator function; it returns a value
expressed in Equation 6 as follows: 𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1,𝑑𝑑𝑑𝑑 = ∑𝑖𝑖𝑖𝑖=0 ∑𝑖𝑖𝑖𝑖=0 ∑𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙, X 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+is1,true
𝐻𝐻𝐻𝐻 𝐵𝐵𝐵𝐵 of 1
𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙
if its argument , and a value of 06otherwise.
𝑖𝑖𝑖𝑖 𝑑𝑑𝑑𝑑
𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙 Equation 9 is obtained
𝑖𝑖𝑖𝑖 as:
𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1,𝑑𝑑𝑑𝑑 = ∑𝐻𝐻𝐻𝐻 𝐵𝐵𝐵𝐵
𝑖𝑖𝑖𝑖=0 ∑𝑖𝑖𝑖𝑖=0 ∑𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙, X 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1, 𝑑𝑑𝑑𝑑 , (6) 6
𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
0≤ D = 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+10,≤for 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1
satisfying 0 ≤� 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 �𝑙𝑙𝑙𝑙 � + 1, <𝑙𝑙𝑙𝑙 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 < 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 − 𝐵𝐵𝐵𝐵 (9)+ 1 =
0 ≤ d, < D = D l+1 for i l+1,d,j l+1
< satisfying i l+1 𝑖𝑖𝑖𝑖< Hl ,–𝑗𝑗𝑗𝑗 H +1, 𝑙𝑙𝑙𝑙
� <= 𝐻𝐻𝐻𝐻 � 2𝜕𝜕𝜕𝜕−𝑙𝑙𝑙𝑙 𝐻𝐻𝐻𝐻
𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 if 𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 > 0, 9
2𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑
0 ≤ d, < D = 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 , for 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 , 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 satisfying 0 ≤ 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 < 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 − 𝐻𝐻𝐻𝐻 + 1, < 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 < 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 − 𝐵𝐵𝐵𝐵 + 1 = 0
< j l+1 < B l – B +𝐵𝐵𝐵𝐵1𝑙𝑙𝑙𝑙+1 = B l+1ininthis equation, 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙𝑖𝑖𝑖𝑖+1,𝑑𝑑𝑑𝑑𝑙𝑙𝑙𝑙 refers
thisequation, refers to the element of 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 indexed 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙+1 + 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 + 𝑗𝑗𝑗𝑗, 𝑑𝑑𝑑𝑑 𝑙𝑙𝑙𝑙 .
𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙
in this equation,
to the element 𝑙𝑙𝑙𝑙 refers
𝑥𝑥𝑥𝑥 𝑖𝑖𝑖𝑖+1,𝑑𝑑𝑑𝑑of x l indexed to the x l+1element
+ j l+1 + of j, d𝑥𝑥𝑥𝑥l . indexed 𝑥𝑥𝑥𝑥 where: + 𝑗𝑗𝑗𝑗 y –+the 𝑗𝑗𝑗𝑗, 𝑑𝑑𝑑𝑑alias . for x . ReLu’s objective is to make
l+1
𝑖𝑖𝑖𝑖
the CNN more nonlinear. Since the semantic content
ReLu layer of an image (in this case, a tomato fruit) is a highly
A multi-layer neural network’s ReLu layer is a nonlin- nonlinear mapping of input pixel values.
ear activation function. In this layer, all weak values, 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙
Batch normalization 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1 layer
= 𝐻𝐻𝐻𝐻 , 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1 = 𝐵𝐵𝐵𝐵 , 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 = 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙 10
and the processed𝑦𝑦𝑦𝑦𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 image= are eliminated
max�0, 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 and
�, switched 7
𝑙𝑙𝑙𝑙 𝑖𝑖𝑖𝑖.𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 Normalization of the inputs to a layer is done in tiny
𝑦𝑦𝑦𝑦𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 = max�0,
out𝑥𝑥𝑥𝑥for �,
𝑖𝑖𝑖𝑖.𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑zeros. Only when the node’s input exceeds where: 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+17, and 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1
𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1 batches using a 𝑙𝑙𝑙𝑙 deep-learning training technique
a predetermined0threshold < i < 𝐻𝐻𝐻𝐻 , 0 < 𝐽𝐽𝐽𝐽 is< activated.
this feature 𝐵𝐵𝐵𝐵 = 𝐵𝐵𝐵𝐵 , 𝜕𝜕𝜕𝜕𝑎𝑎𝑎𝑎𝑑𝑑𝑑𝑑 0 ≤ 𝑑𝑑𝑑𝑑 ≤ 𝐷𝐷𝐷𝐷 ≤ 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 .
𝑙𝑙𝑙𝑙
𝑚𝑚𝑚𝑚𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 is accelerated
0 < i < 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1 , 0 < 𝐽𝐽𝐽𝐽 < 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 = 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1 , 𝜕𝜕𝜕𝜕𝑎𝑎𝑎𝑎𝑑𝑑𝑑𝑑 0 ≤ 𝑑𝑑𝑑𝑑 ≤ 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙 ≤ 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 . known as batch normalization. Training
The output will be zero if the input is less than zero. max: 𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 = 0 < 𝑖𝑖𝑖𝑖 < 𝐻𝐻𝐻𝐻, 0 < 𝑗𝑗𝑗𝑗 < 𝐵𝐵𝐵𝐵 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 × 𝐻𝐻𝐻𝐻 + 𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 × 𝐵𝐵𝐵𝐵
However, the input becomes linearly related to the de- by the 𝑖𝑖𝑖𝑖batch normalization
,𝑑𝑑𝑑𝑑 layer, which also lessens
𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1

once it exceeds a𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕 𝑙𝑙𝑙𝑙where: 0 the impact of𝑙𝑙𝑙𝑙+1

𝑙𝑙𝑙𝑙+1 initialization 𝑙𝑙𝑙𝑙+1 after 𝑙𝑙𝑙𝑙+1
< 𝐵𝐵𝐵𝐵 the convolution oper- 𝑙𝑙𝑙𝑙+1
pendent
𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑 variable certain = threshold. > 0�,≤ 𝑖𝑖𝑖𝑖 < 𝐻𝐻𝐻𝐻 , 0 ≤ 𝑗𝑗𝑗𝑗 8 , and 0≤ 𝑑𝑑𝑑𝑑 < 𝐷𝐷𝐷𝐷 –
𝑖𝑖𝑖𝑖,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑
𝑙𝑙𝑙𝑙 �𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑
This𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕implies
𝑙𝑙𝑙𝑙 = �𝑥𝑥𝑥𝑥 > 0�,
that the deep neural network’s training
𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 𝑑𝑑𝑑𝑑𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖𝑙𝑙𝑙𝑙,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑 8
ation. This layer is added before the input of each con-
𝑖𝑖𝑖𝑖,𝑗𝑗𝑗𝑗,𝑑𝑑𝑑𝑑
data set can be processed volutional layer to ensure that each layer’s input has the
ere ‖. ‖ – the vation functions. where ‖. ‖ –more
the quickly than other acti- same distributions and to reduce inner covariate shifts
This prevents the sum from reaching
zero. The ReLu layer can be considered as independent during training. Therefore, the process of learning is
filtering for each element in the input: The input size is regulated and a deep network can be trained with sig-
unchanged by a ReLu layer. ReLu layers maintain the nificantly fewer training epochs.
input’s original size, so/and y have the same size. This
is expressed in Equation 7 as: MaxPooling layer
Two input arguments are required by a MaxPool layer,
yi,j,d = max{0, x li, j,d}, (7) height, and core width. The kernel moves across the
0 < i < H l+1, 0 < j < Bl = B l+1n and 0 ≤ d ≤ D l ≤ D l+1. pixels in a straight line at the predetermined step size,
 Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 411
9-11
9-11
beginning in the upper left corner of the feature map.
The kernel window’s highest-valued pixel serves as
Softmax layer
It is mainly used to represent certainty probabilities in
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕the source�of 𝑙𝑙𝑙𝑙the � value for𝑙𝑙𝑙𝑙 the associated node in the neural network outputs by scaling output between 0
� 𝑙𝑙𝑙𝑙� = � 2𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 if 𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 > 0, 9
2𝜕𝜕𝜕𝜕pooling
𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 layer. Max CNNs’ pooling layers are a crucial and 1. It is possible to compute normalization by divid-
0
component. They keep the number of network param- ing the output under the study’s exp value by the sum
eters to a minimum while summarizing the activation of all possible outputs’ exp values. The softmax func-
maps, the result of pooling 𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 (y, or x in its place), (H × B tion transforms a vector of k real values into a vector of
× D) will �be𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕the size of � H𝑙𝑙𝑙𝑙l+1 � + B l+1 +𝑙𝑙𝑙𝑙 D l+1 in order 3-ten- k real values with a total of 1. Softmax converts input
� =� 2𝜕𝜕𝜕𝜕 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 if 𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 > 0, 9
2𝜕𝜕𝜕𝜕 𝑙𝑙𝑙𝑙 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑
𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
sor as seen in Equation 10: 0 values, which can be either positive or negative, zero
� or higher than 1, into values between 0 and 1, mak-
2𝜕𝜕𝜕𝜕 𝑙𝑙𝑙𝑙 𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 if 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙
𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑑𝑑𝑑𝑑 > 0, 𝑙𝑙𝑙𝑙+1 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙+1
9𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙 (10)
0 𝐻𝐻𝐻𝐻 = , 𝐵𝐵𝐵𝐵 = , 𝐷𝐷𝐷𝐷 = 𝐷𝐷𝐷𝐷 10 them understandable as probabilities. Any small
ing
𝐻𝐻𝐻𝐻 𝐵𝐵𝐵𝐵
or negative, input is transformed into a small chance
e: 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1 , and 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+where: 1 H l+1, and D l+1 – the input to the pooling layer. by Softmax, and any large input is transformed into
A max-pooling operator maps
𝑚𝑚𝑚𝑚𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕
𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙 a subregion 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 to its maxi- a large probability.
𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1 = , 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1 = in , 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙
= 𝐷𝐷𝐷𝐷11 𝑙𝑙𝑙𝑙+1 10
max: 𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 mum value. Using
= 0 < 𝑖𝑖𝑖𝑖 < 𝐻𝐻𝐻𝐻, 0 < 𝑗𝑗𝑗𝑗 < 𝐵𝐵𝐵𝐵 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙exact mathematics
𝐻𝐻𝐻𝐻 × 𝐻𝐻𝐻𝐻 + 𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗 𝐵𝐵𝐵𝐵 Equation
× 𝐵𝐵𝐵𝐵 + 𝑗𝑗𝑗𝑗, 𝑑𝑑𝑑𝑑. 11
𝑖𝑖𝑖𝑖 ,𝑑𝑑𝑑𝑑 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1
𝑙𝑙𝑙𝑙
where: 𝐻𝐻𝐻𝐻𝑙𝑙𝑙𝑙as: 𝑙𝑙𝑙𝑙+1
, and 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 Output layer
e:
𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1
= 0𝐻𝐻𝐻𝐻≤, 𝐵𝐵𝐵𝐵𝑖𝑖𝑖𝑖𝑙𝑙𝑙𝑙+1 <
𝐵𝐵𝐵𝐵 𝑙𝑙𝑙𝑙+1
= 𝐻𝐻𝐻𝐻𝐵𝐵𝐵𝐵 , 𝐷𝐷𝐷𝐷, 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1 𝑙𝑙𝑙𝑙+1
0 ≤=𝑗𝑗𝑗𝑗 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙 < 𝐵𝐵𝐵𝐵 10, and𝑚𝑚𝑚𝑚𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 0≤ 𝑑𝑑𝑑𝑑 < 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1
– The neural network’s output layer is the final layer that
max:max: 𝑦𝑦𝑦𝑦 𝑙𝑙𝑙𝑙+1,𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 = 0 < 𝑖𝑖𝑖𝑖 < 𝐻𝐻𝐻𝐻, 0 < 𝑗𝑗𝑗𝑗 < 𝐵𝐵𝐵𝐵 𝑥𝑥𝑥𝑥 𝑙𝑙𝑙𝑙 𝐻𝐻𝐻𝐻 + 𝑖𝑖𝑖𝑖, produces
×(11) the
𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 × 𝐵𝐵𝐵𝐵 + desired predictions.
𝑗𝑗𝑗𝑗, 𝑑𝑑𝑑𝑑. 11 An output layer in
𝑖𝑖𝑖𝑖 ,𝑑𝑑𝑑𝑑 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 a neural network generates the desired outcome pre-
where: 0 ≤ 𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙+1 < 𝐻𝐻𝐻𝐻 𝑙𝑙𝑙𝑙+1 , 0 ≤ 𝑗𝑗𝑗𝑗𝑙𝑙𝑙𝑙+1 < 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1 , and 0≤ 𝑑𝑑𝑑𝑑 < 𝐷𝐷𝐷𝐷 𝑙𝑙𝑙𝑙+1 – diction. Before determining the final output, it applies
𝑚𝑚𝑚𝑚𝜕𝜕𝜕𝜕𝜕𝜕𝜕𝜕 𝑙𝑙𝑙𝑙+1
< 𝑖𝑖𝑖𝑖 < 𝐻𝐻𝐻𝐻, 0 < 𝑗𝑗𝑗𝑗 < 𝐵𝐵𝐵𝐵 𝑥𝑥𝑥𝑥𝑖𝑖𝑖𝑖𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙+1 × 𝐻𝐻𝐻𝐻 + 𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗 × 𝐵𝐵𝐵𝐵 + 𝑗𝑗𝑗𝑗, 𝑑𝑑𝑑𝑑. 11 its own set of weights and biases. For some issues, the
hidden layer activation function may be different from
𝑙𝑙𝑙𝑙+1
< 𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙+1 , andwhere: 0≤ 𝑑𝑑𝑑𝑑 < 0 ≤𝐷𝐷𝐷𝐷i𝑙𝑙𝑙𝑙+1
l+1
<–H l+1, 0 ≤ j l+1 < B l+1 and 0 ≤ d < D l+1 – the output layer activation function. For instance, in
depicts a pooling local operator with straight forward classification issues, the final classes are derived using
computation. softmax activations. A summary of the proposed CNN
Fully connected layer Architecture is given in Figure 7.

The flat layer, a two-dimensional (2d) layer, provides

input to a completely connected layer. The affine Feature extraction
function receives data from the smoothing layer be- Different subsections of the feature extraction phase,
fore passing it on to the nonlinear function. One FC such as color extraction, texture extraction, and in-
(fully connected) is the result of combining one affine fected area extraction were considered. An RGB im-
function with one nonlinear function. The suggested age, in terms of conventional color space, is a mixture
architecture’s final probability for each label is pro- of the colors red, green, and blue. Therefore, it is ap-
vided in this layer. The FC layer transforms incom- propriate to determine the proportion of red, green,
ing neurons using weighted linear transformations and blue components in each image pixel. One of the
and sends outputs to nonlinear neurons. Every neu- main components for detecting tomato disease is color
ron in one layer is connected to every neuron in the extraction. This technique takes a sample image of
other layers, forming a succession of fully connected a tomato and extracts a spectrum of colors from light-
layers that make up a fully connected neural network. est to darkest. Entropy calculation is used to perform
Fully connected networks have the major benefit of
being “structure-agnostic”, meaning that no specific
12-14 texture detection on segmented captured images. This
technique first converts the segmented image to gray-
12-14
assumptions about the input are required. The de-
scriptions of mathematical models with completely
scale. Following grayscale conversion, the system uses
a conventional entropy calculation method to extract
connected layers are provided in Equation 12. Con-
12-14
sidering a convolution or pooling layer/node with the
the local entropy of the image. The system determines
the minimum and maximum pixels present in the lo-
dimensions: cal entropy matrix following its extraction. The texture
quality coefficient will be equal to the mean of the cal-
[𝑖𝑖𝑖𝑖] 𝑛𝑛𝑛𝑛(𝑖𝑖𝑖𝑖−𝑙𝑙𝑙𝑙) (𝑖𝑖𝑖𝑖−1) [] [] []
𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 = ∑(𝑙𝑙𝑙𝑙−1) 𝐵𝐵𝐵𝐵𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 1𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 + 𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙 → 𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 = 𝜑𝜑𝜑𝜑 [𝑖𝑖𝑖𝑖] �𝑧𝑧𝑧𝑧𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖culated
�. minimum and 12 maximum pixel range from the
(12) local entropy matrix. A rough idea of tomato texture
[𝑙𝑙𝑙𝑙] [𝑖𝑖𝑖𝑖] [𝑖𝑖𝑖𝑖]
∑𝑛𝑛𝑛𝑛(𝑖𝑖𝑖𝑖−𝑙𝑙𝑙𝑙)
(𝑙𝑙𝑙𝑙−1) 𝑖𝑖𝑖𝑖𝐵𝐵𝐵𝐵 𝑖𝑖𝑖𝑖
1𝐵𝐵𝐵𝐵
(𝑖𝑖𝑖𝑖−1)
𝑙𝑙𝑙𝑙 + 𝑑𝑑𝑑𝑑 𝑖𝑖𝑖𝑖 → 𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 = 𝜑𝜑𝜑𝜑 [𝑖𝑖𝑖𝑖]
�𝑧𝑧𝑧𝑧 𝑖𝑖𝑖𝑖 �. 12 can be obtained from the texture quality coefficient
𝐴𝐴𝐴𝐴 (𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = ∑𝑖𝑖𝑖𝑖=1 ∑𝑖𝑖𝑖𝑖=1 ∑𝑘𝑘𝑘𝑘=1 𝐵𝐵𝐵𝐵 (𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) × 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙2𝐵𝐵𝐵𝐵(calculated from the local 13 entropy matrix. The creat-
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛
𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑘𝑘𝑘𝑘),
(𝑖𝑖𝑖𝑖−1) [] [𝑖𝑖𝑖𝑖] [𝑖𝑖𝑖𝑖] ed system is trained using an extensive set of texture
+ 𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙 → 𝜕𝜕𝜕𝜕 The
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖 𝑛𝑛𝑛𝑛
= input
𝜑𝜑𝜑𝜑 [𝑖𝑖𝑖𝑖]
�𝑧𝑧𝑧𝑧 𝑖𝑖𝑖𝑖 is
�. the output of the 12
pooling𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 opera-
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛
quality coefficients that were extracted from various
𝑧𝑧𝑧𝑧) = ∑𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=1 ∑𝑖𝑖𝑖𝑖=1tion ∑𝑘𝑘𝑘𝑘=1with𝐵𝐵𝐵𝐵H(l+1 𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗,+𝑘𝑘𝑘𝑘B)l+1×+D𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙l+1𝑙𝑙𝑙𝑙.2𝐵𝐵𝐵𝐵(𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑘𝑘𝑘𝑘( ), 13 𝑎𝑎𝑎𝑎 (𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) �
𝐵𝐵𝐵𝐵 𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = � � �tomato sample 2,
𝑚𝑚𝑚𝑚images in the created dataset as:
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=1 𝑖𝑖𝑖𝑖=1 𝑘𝑘𝑘𝑘=1
∑𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 ∑𝑛𝑛𝑛𝑛𝑘𝑘𝑘𝑘=1 𝐵𝐵𝐵𝐵 (𝑖𝑖𝑖𝑖,𝐵𝐵𝐵𝐵𝑗𝑗𝑗𝑗,(𝑘𝑘𝑘𝑘𝑥𝑥𝑥𝑥,) 𝑦𝑦𝑦𝑦,×𝑧𝑧𝑧𝑧)𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙=𝑙𝑙𝑙𝑙2𝐵𝐵𝐵𝐵 𝑎𝑎𝑎𝑎 (𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) � 13
𝑖𝑖𝑖𝑖=1 � (𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑘𝑘𝑘𝑘�
), � 𝑚𝑚𝑚𝑚2 ,
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=1 𝑖𝑖𝑖𝑖=1 𝑘𝑘𝑘𝑘=1
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛
 412 Journal of Plant Protection Research 63 (4), 2023

12-14

[𝑖𝑖𝑖𝑖] 𝑛𝑛𝑛𝑛(𝑖𝑖𝑖𝑖−𝑙𝑙𝑙𝑙) (𝑖𝑖𝑖𝑖−1) [] [] []

𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 = ∑(𝑙𝑙𝑙𝑙−1) 𝐵𝐵𝐵𝐵𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 1𝐵𝐵𝐵𝐵𝑙𝑙𝑙𝑙 + 𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 𝑙𝑙𝑙𝑙 → 𝜕𝜕𝜕𝜕𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 = 𝜑𝜑𝜑𝜑 [𝑖𝑖𝑖𝑖] �𝑧𝑧𝑧𝑧𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 �.

𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛
𝐴𝐴𝐴𝐴 (𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = ∑𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛
𝑖𝑖𝑖𝑖=1 ∑𝑖𝑖𝑖𝑖=1 ∑𝑘𝑘𝑘𝑘=1 𝐵𝐵𝐵𝐵 (𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) × 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙2𝐵𝐵𝐵𝐵(𝑖𝑖𝑖𝑖,𝑖𝑖𝑖𝑖,𝑘𝑘𝑘𝑘),

𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛

𝑎𝑎𝑎𝑎 (𝑖𝑖𝑖𝑖,
𝐵𝐵𝐵𝐵(𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = � � �
𝑖𝑖𝑖𝑖=1 𝑖𝑖𝑖𝑖=1 𝑘𝑘𝑘𝑘=1
Fig. 7. Summary of the proposed backbone structure

13

𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑗𝑗𝑗𝑗=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛

𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 (14)
𝐴𝐴𝐴𝐴(𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = � � � 𝐵𝐵𝐵𝐵(𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) × log 2𝐵𝐵𝐵𝐵(𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ), 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 = �𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝐹𝐹𝐹𝐹𝑇𝑇𝑇𝑇, 14
𝑖𝑖𝑖𝑖=1 𝑖𝑖𝑖𝑖=1 𝑘𝑘𝑘𝑘=1
where: Tp – the number of positive samples that were
𝑗𝑗𝑗𝑗=𝑛𝑛𝑛𝑛
𝑖𝑖𝑖𝑖=𝑛𝑛𝑛𝑛 𝑛𝑛𝑛𝑛
𝑛𝑛𝑛𝑛 (𝑖𝑖𝑖𝑖, 𝑗𝑗𝑗𝑗, 𝑘𝑘𝑘𝑘 ) � (13) correctly classified, that is, the real positive samples
𝐵𝐵𝐵𝐵(𝑥𝑥𝑥𝑥, 𝑦𝑦𝑦𝑦, 𝑧𝑧𝑧𝑧) = � � � 𝑚𝑚𝑚𝑚2 , that the classifier also correctly classified as positive
𝑖𝑖𝑖𝑖=1 𝑖𝑖𝑖𝑖=1 𝑘𝑘𝑘𝑘=1
samples, and Fp – the number of negative samples that
where: B(x,y,z) – the calculated probability index ma- were incorrectly classified as positive, that is, the real
trix, and m2 – the size of the segmented image. negative samples that the classifier incorrectly classi-
fied as positive samples.

Mean average precision (mAP)

Results, Discussion and Conclusions To determine the mAP, the mean15-18 average precision
compares the detected box to the ground-truth bound-
The specifics of the experiments including param- ing box. When the ratings are high, the model’s detec-
eter fitting and the outcomes of the experiments are tions are more accurate. Average precision is defined
covered in this section. Each model’s performance is for datasets with multiple classes as in Equation 15
clearly presented. as:

∑𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖� (15)

Quantitative Analysis 𝑚𝑚𝑚𝑚𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 = 𝑁𝑁𝑁𝑁 ′ , 15

where: Σ where:
The primary indicators for evaluating a deep learning ∑ – represents
– represents the sum the
of averagesum of average precision precision
and 𝑁𝑁𝑁𝑁 ′and
– represents the tota
model’s detection performance are precision, average N’– represents the total number of all classes in the
classes in the dataset.
precision, accuracy, and recall. The training time and dataset.
image detection time were used to obtain the mean
average precision (mAP) of each model’s performance Accuracy
in this study. Accuracy, or the percentage 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 of correctly detected im-
𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = �𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝐹𝐹𝐹𝐹𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 + 𝐹𝐹𝐹𝐹𝑎𝑎𝑎𝑎, 16
ages in all instances, is a term used to describe the rate
Precision at which images are successfully detected. Addition-
ally, the proportion of samples correctly identified
Precision is the ratio of the number of positively de- for a given test 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 database= to �∑the (𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖total )2�
− 𝑂𝑂𝑂𝑂𝑖𝑖𝑖𝑖number of data 17
𝑎𝑎𝑎𝑎 ,
tected samples that were correctly identified as positive successfully classified by the predictor is proportional
to the number of positively detected samples. Preci- to the total amount of data. Accuracy is described in
sion is defined in Equation 14 as:
𝐼𝐼𝐼𝐼𝑂𝑂𝑂𝑂𝐼𝐼𝐼𝐼 Equation 16: 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 �
= 𝐼𝐼𝐼𝐼𝑎𝑎𝑎𝑎𝐼𝐼𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐼𝐼𝐼𝐼𝑖𝑖𝑖𝑖 𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃. 18
 𝑁𝑁𝑁𝑁
where: Σ – represents the sum of average precision and 𝑁𝑁𝑁𝑁 ′ – represents the total number of all
classes in the dataset.
Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 413
15-18
The outcome of the experimental results showed
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎
𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = �𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝐹𝐹𝐹𝐹𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 + 𝐹𝐹𝐹𝐹𝑎𝑎𝑎𝑎, that the proposed CNN backbone with single shot
16
(16)
detector algorithm achieved excellent results of 0.991
precision, 0.994 mean average precision, 0.988 accu-
where: Tp – the number of positive samples that were
– (the 𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 − 𝑂𝑂𝑂𝑂𝑖𝑖𝑖𝑖 ) � of negative sam- racy, and a root mean square error of 1.21 The per-
2
correctly classified, 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 =Fp �∑ number
𝑎𝑎𝑎𝑎 , 17 formance evaluation of the proposed CNN backbone
ples that were incorrect, Tn – negative classes that were
∑𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖Fn with SSD Algorithm is presented in Table 1.
correctly 𝑚𝑚𝑚𝑚𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 predicted
= as negative, and �𝑁𝑁𝑁𝑁–′ , positive class- 15
The accuracy, precision, mean average precision
𝐼𝐼𝐼𝐼𝑂𝑂𝑂𝑂𝐼𝐼𝐼𝐼Σ = es that were
𝐼𝐼𝐼𝐼𝑎𝑎𝑎𝑎𝐼𝐼𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐼𝐼𝐼𝐼𝑖𝑖𝑖𝑖 falsely
𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 predicted as negative.
�𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 18
where: – represents the sum of average precision
𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃. and 𝑁𝑁𝑁𝑁 ′
– represents (mAP),
the totalintersection
number of over
all union (IOU), and root mean
square error (RMSE), at 1000th for the proposed
classes in the dataset. Root mean square error (RMSE)
model were recorded. The performance of the pro-
The average distance between the values in the data-
set and those predicted by the model. A given model
15-18 posed model was compared with existing models like
ResNet-50, AlexNet, VGG16, and VGG19. The model’s
can “fit” a dataset more accurately when the RMSE is
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 +square 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 performance is shown in Table 2.
low.𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃
The root =means �𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇error or RMSE is, calcu- 16
+ 𝐹𝐹𝐹𝐹𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 + 𝐹𝐹𝐹𝐹𝑎𝑎𝑎𝑎 The classes in the dataset included black mold
lated using the following formula:
diseases, radial ring diseases, spotted wilt diseases,
spotted wilt viruses, and healthy tomato fruits. The
(𝑃𝑃𝑃𝑃 − 𝑂𝑂𝑂𝑂𝑖𝑖𝑖𝑖 ) � 2 performance of the proposed model in each class
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = �𝑚𝑚𝑚𝑚𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴 ∑ 𝑖𝑖𝑖𝑖 = ∑𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 𝑖𝑖𝑖𝑖(17)
𝑎𝑎𝑎𝑎 , 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖17 �𝑁𝑁𝑁𝑁 ′ , was evaluated 15 and the outcome is seen in Table 3.
The results of the performance evaluation show that
where: where: Σ –Σthe – represents
sum of RMSE, the sum Pi is ofthe averagepredicted precision valueand 𝑁𝑁𝑁𝑁 ′ – represents the total number of all
the proposed model obtained higher accuracy, pre-
𝐼𝐼𝐼𝐼𝑂𝑂𝑂𝑂𝐼𝐼𝐼𝐼 = 𝐼𝐼𝐼𝐼𝑎𝑎𝑎𝑎for the ith𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖in
𝐼𝐼𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐼𝐼𝐼𝐼𝑖𝑖𝑖𝑖
classes –𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃
observation
the dataset. �𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 in the dataset,18Oi – the ob-
𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃. cision, and mean average precision on healthy to-
served value for the observation in the dataset, and is
mato fruit than black mold, spotted wilt, and radial
the sample size.
ring.
Figure 8 shows sample test results from the experi-
Intersection over union (IOU) 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎�
𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = 𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 + 𝐹𝐹𝐹𝐹𝑇𝑇𝑇𝑇 + 𝑇𝑇𝑇𝑇𝑎𝑎𝑎𝑎 +ments.
𝐹𝐹𝐹𝐹𝑎𝑎𝑎𝑎 , The results
16 show the class name/ the disease of
A metric quantifies how well the predicted and actual
the tomato fruit and the percentage of the detection re-
boxes match. By dividing the area of the intersection
sults. Figure 9 shows the train accuracy and loss graphs
between the two boxes by the area of their union,2 the
(𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 − 𝑂𝑂𝑂𝑂𝑖𝑖𝑖𝑖 ) � of the models used in this study.
IoU is determined using the following
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = �∑equation. The 𝑎𝑎𝑎𝑎 , 17
To identify the model that was the most effective at
accuracy of the prediction increases with IoU.
detecting tomato fruit diseases accuracy, mean average
𝐼𝐼𝐼𝐼𝑂𝑂𝑂𝑂𝐼𝐼𝐼𝐼 = 𝐼𝐼𝐼𝐼𝑎𝑎𝑎𝑎𝐼𝐼𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐼𝐼𝐼𝐼𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 �𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼 𝐴𝐴𝐴𝐴𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃. (18) 18
precision, precision, recall, and root mean square error
score values were used as evaluation metrics to assess
A novel convolutional neural network method was
proposed in this study for tomato disease detection.
The proposed model serves as a backbone for the sin-
gle shot object detector algorithm. The main aim of this Table 1. Performance evaluation of the proposed backbone-SSD
research was to improve the accuracy of the existing Metrics Results
deep learning models in plant disease detection. The Precision 0.991
total epochs for training and testing were 100 epochs
Mean average precision (mAP) 0.994
at 1000 iterations. Seventy percent of the total datasets
Accuracy 0.988
was utilized for training, and 30% was used for data
testing. Table 1 shows the results of the proposed CNN Recall 0.991

with SDD in detecting tomato fruit disease. Root mean square error 1.21

Table 2. Models performance evaluation on test data

Models Accuracy Precision mAP Recall RMSE

SSD-ResNet-50 0.985 0.990 0.995 0.991 1.27
SSD-AlexNet 0.963 0.953 0.952 0.981 1.28
SSD-VGG-16 0.963 0.951 0.953 0.97 1.28
SSD-VGG-19 0.966 0.954 0.962 0.98 1.28
SSD-Proposed CNN 0.988 0.991 0.994 0.991 1.21
mAP – mean average precision
414 Journal of Plant Protection Research 63 (4), 2023

Table 3. Proposed model performance in each class A crop conditional model was developed which uti-
Class Accuracy Precision mAP lized a unique CNN design in conjunction with crop
Healthy 0.985 0.985 0.982 metadata to recognize 17 diverse diseases in five crops.
These crops included rapeseed, barley, rice, wheat, and
Black mold 0.975 0.974 0.971
corn. From the obtained robust features of a large-size
Spotted wilt 0.953 0.951 0.950
multi-crop dataset, the model quickly learned similar
Radial ring 0.967 0.966 0.961
disease symptoms in different crops, which decreased
mAP – mean average precision
the complexity of the classification function (Picon
et al. 2019)
Durmuş et al. (2017) proposed a DL method for
the performance of the pre-trained models. A graph tomato leaf disease detection. They aimed at using
showing the validation accuracy for the pre-trained a robot to run real-time plant leaf detection manually
models was created using the validation accuracy cal- or autonomously in the field or greenhouse. Train-
culated in Figure 9. ing and validation were done by adopting AlexNet
and Squeeze models on the plant image dataset. The
examined tomato leaf diseases in their research cause
Discussions a physical change. RGB cameras can observe these al-
terations in the leaves. The output of their novel ap-
proach showed that the AlexNet model performed
This section presents a thorough analysis of recent slightly better than the SqueezeNet model. They
research on plant disease detection via deep learn- concluded that the SqueezeNet model is 80 times
ing methods. Ignoring the early indicators of plant smaller than the AlexNet Model, and the cause of
disease in the agricultural sector can result in losses the differences was attributed to the Caffe format
in food harvests and ultimately lead to the collapse of (Durmuş et al. 2017).
the global food industry. In a few selected diseases and (Iqbal et al. 2021) proposed a gray level co-oc-
crops, the recent trend of using different machine- currence matrix (GLCM) algorithm to calculate
learning algorithms for plant disease detection has 13 distinctive statistical features of tomato leaves. To
yielded encouraging results. categorize data, the support vector machine (SVM)

Fig. 8. Sample test results. (A) Sample detected Spotted wilt virus at 0.92%. (B) Sample detected black mold virus at 0.58 and
0.64 respectively. (C) Sample detected radial ring virus at 0.87%. (D) Sample detected Healthy tomato fruit at 0.92%.
 Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 415

A B

Fig. 9. Training-accuracy graphs for the models – A; Training-loss graphs for the models – B

was utilized. The features obtained from the GLCM

higher features of the affected area of the crop to en-
algorithm were implemented as a mobile application.
able accurate disease detection. A 15-layer convolu-
The processed leaf was compared with the features
tional neural network was proposed as the backbone
stored to recognize the tomato leaf disease. The ex-
of a single short shot detector model to address these
perimental findings of their method provide 100% ac-
challenges. This method can identify the affected re-
curacy for healthy leaves, 95% for early blight, 90% for
gion and the shapes of the infected areas.
Septoria leaf, and 85% for late blight (Iqbal et al. 2021).
This research focused on improving the single shot
For effective plant disease identification, pre-
detecting algorithm by proposing a CNN backbone for
trained models based on convolutional neural net-
tomato fruit disease detection. The main aim of this
works (CNN) were used. The focus was on fine-tun-
research was to identify the common disease that af-
ing the hyperparameters of well-known pre-trained
fects the tomato fruit, gather a tomato fruit dataset,
models like DenseNet-121, ResNet-50, VGG-16, and
and enhance the deep learning framework for detect-
Inception V4, in particular. The experiments used the
ing plant disease. The proposed method forecasts and
well-known Plant Village dataset, which contains
creates a new prototype that will offer improved plant
54,305 images of various plant disease species organ-
disease detection performance with less computation-
ized into 38 classes. Through classification accuracy,
al resources in a short time. The results obtained from
sensitivity, specificity, and F1 score, the model’s perfor-
the proposed method performed better than SSD with
mance was assessed. Additionally, a comparison with
ResNet-50, AlexNet, VGG16, and VGG19 backbones
comparable cutting-edge studies was done. The tests
in tomato fruit disease detection. In a subsequent
showed that DenseNet-121 outperformed cutting-
study, we will concentrate on how to enhance the to-
edge models by 99.81% in terms of classification ac-
mato disease dataset to befit general target detection
curacy (Andrew et al. 2022).
algorithms.
To identify plant species in images, most of the
The detection and management of crop and plant
studies reviewed adopted pre-trained CNN models
infestations have been significantly improved by deep
by fine-tuning, conducting computational statistics of
learning technologies. Complex disease and pest iden-
the tomato leaf features, and using a robot in detect-
tification has become possible due to advanced de-
ing tomato plant disease in the greenhouse. The per-
velopments of object detectors. Large and real-time
formance of their methods was compared with other
datasets are supported, automatic feature extraction
state-of-the-art models. Their findings showed that
functionality is also available, and the overall execu-
deep convolutional neural networks perform well in
tion time is decreased. Therefore, deep learning tech-
the identification of plant diseases. However, more
niques can be taken into account for upcoming re-
work is required to improve the previous studies, such
search in the agriculture sector, such as roots, land,
as proposing crop disease models that can extract
weeds, leaves, and fruits for disease identification.
416 Journal of Plant Protection Research 63 (4), 2023

However, the majority of this research is lab-based and Gatahi D. 2020. Challenges and opportunities in tomato pro-
duction chain and sustainable standards introduction.
heavily depends on gathered images of plant diseases International Journal of Horticulture Science and Tech-
and pests. To increase the robustness and generaliza- nology 7 (3): 235–262. DOI: https://doi.org/10.22059/
tion of the deep learning models, future research will ijhst.2020.300818.361
focus on collecting images from various plant growth Ghazal T. 2022. Convolutional neural network based intelli-
gent handwritten document recognition. Computers, Ma-
stages, seasons, and geographic locations. terials & Continua 70 (3): 4563–4581. DOI: https://doi.
Early detection of plant diseases and pests is es- org/10.32604/cmc.2022.021102
sential for halting and controlling their growth., Ac- Golan K., Kot I., Kmieć K., Górska-Drabik E. 2023. Approaches
to integrated pest management in Orchards: Comstockaspis
curate identification and prediction require the incor-
perniciosa (comstock) case study. Mdpi- Agriculture 13 (1):
poration of meteorological and plant health data, such 131. DOI: https://doi.org/10.3390/agriculture13010131
as temperature and humidity. Deep learning model Guravaiah K., Bhavadeesh Y.S., Shwejan P., Vardhan A.H., La-
training and network learning can benefit from unsu- vanya S. 2023. Third eye: object recognition and speech
generation for visually impaired. Procedia Computer
pervised learning as well as combining prior knowl- Science 218: 1144–1155. DOI: https://doi.org/10.1016/j.procs.
edge of brain-like computers with human visual cog- 2023.01.093
nition. Collaboration between experts in agriculture Haar L. V., Elvira T., Ochoa O. 2023. An analysis of explainabil-
ity methods for convolutional neural networks. Engineer-
and plant protection in future research is necessary to
ing Applications of Artificial Intelligence 117: 105606. DOI:
realize the full potential of this technology. Their skills https://doi.org/10.1016/j.engappai.2022.105606
and knowledge must be combined with deep learning Hemathilake D., Gunathilake D. 2022. Agricultural productivity
algorithms and models, and the resulting information and food supply to meet increased demands. Future Foods
2022: 539–553. DOI: https://doi.org/10.1016/B978-0-323-
will be incorporated into farming machinery. 91001-9.00016-5
Hofman-Bergholm M. 2023. A transition towards a food and
agricultural system that includes both food security and
planetary health. Mdpi-Foods 12 (1): 12. DOI: https://doi.
References
org/10.3390/foods12010012
Humbal A., Pathak B. 2023. Application of nanotechnology in
Albattah W., Javed A., Nawaz M., Masood M., Albahli S. 2022. plant growth and diseases management: tool for sustainable
Artificial Intelligence-based drone system for multiclass agriculture. p. 145–168. In: “Agricultural and Environmen-
plant disease detection using an improved efficient con- tal Nanotechnology” (F-L. Fabian, J.K. Patra, eds.). Springer,
volutional neural network. Frontiers in Plant Science 13: Singapore, 674 pp. DOI: https://doi.org/10.1007/978-981-
1003152. DOI: https://doi.org/10.3389/fpls.2022.808380. 19-5454-2_6
Andrew J., Eunice J., Popescu D. E., Chowdary M.K., He- Iqbal N., Mumtaz R., Shafi U., Zaidi S. M. H. 2021. Gray level
manth J. 2022. Deep learning-based leaf disease detection co-occurrence matrix (GLCM) texture based crop clas-
in crops using images for agricultural applications. Agron- sification using low altitude remote sensing platforms.
omy 12 (10): 2395. DOI: https://doi.org/10.3390/agrono- PeerJ Computer Science 7 (8): e536. DOI: https://doi.
my12102395 org/10.7717/peerj-cs.536
Ates C., Fidan H., Karacaoglu M., Dasgan H. 2019. The iden- Ji Y., Liu S., Hao Y. 2023. Realization of convolutional neural
tification of the resistance levels of Fusarium oxysporum network based on FPGA. p. 761–765. In: Proceedings of
f. sp. radicis-lycopersici and tomato yellow leaf curl viruses “Third International Conference on Computer Vision and
in different tomato genotypes with traditional and molecu- Data Mining (ICCVDM 2022). Hulun Buir, China. DOI:
lar methods. Applied Ecology and Environmental Research https://doi.org/10.1117/12.2660023
17 (2): 2203–2218. DOI: http://dx.doi.org/10.15666/aeer Khan H. R., Gillani Z., Jamal M.H., Athar A., Chaudhry M.T.,
Bhujel A., Kim N.-E., Arulmozhi E., Basak J., Kim H.-T. 2022. Chao H., He Y., Chen M. 2023. Early identification of crop
A lightweight attention-based convolutional neural net- type for smallholder farming systems using deep learning
works for tomato leaf disease classification. Mdpi- Agri- on time-series sentinel-2 imagery. Mdpi-Sensors 23 (4):
culture 12 (2): 228. DOI: https://doi.org/10.3390/agricul- 1779. DOI: https://doi.org/10.3390/s23041779
ture12020228 Knapp S., Peralta I.E. 2016. The tomato (Solanum lycopersicum
Bouni M., Hssina B., Douzi K., Douzi S. 2023. Impact of pre- L., Solanaceae) and its botanical relatives. p. 7–21. In “The
trained deep neural networks for tomato leaf disease predic- Tomato Genome” (M. Causse, J. Giovannoni, M. Bouzayen,
tion. Journal of Electrical and Computer Engineering 2023 Z. Mohamed, eds.). Springer Berlin, Heidelberg, 259 pp.
(1): 1–11. DOI: https://doi.org/10.1155/2023/5051005 DOI: https://doi.org/10.1007/978-3-662-53389-5_2
Caruso A.G.S.B., Parrella G, Rizzo R., Davino S., Panno S. Kremneva O., Danilov R., Gasiyan K., Ponomarev A. 2023.
2022. Tomato brown rugose fruit virus: a pathogen that is Spore-trapping device: an efficient tool to manage fungal
changing the tomato production worldwide. Annals of diseases in winter wheat crops. Plants 12 (2): 391. DOI:
Applied Biology 181 (3): 258–274. DOI: https://doi.org/ https://doi.org/10.3390/plants12020391
10.1111/aab.12788 Leite G.L.D., Fialho A. 2018. Sustainable management of ar-
Durmuş H., Güneş E. O., Kırcı M. 2017. Disease detection on thropod pests of tomato. p. 305–311. In: “Sustainable
the leaves of the tomato plants by using deep learning. p. Management of Arthropod Pests of Tomato” (W. Wakil,
1–5. In: Proceedings of “6th International Conference on G.E. Brust, T.M. Perring, eds.). Academic Press. DOI: htt-
Agro-Geoinformatics”. 7–10 August 2017, Fairfax, VA, ps://doi.org/10.1016/B978-0-12-802441-6.00014-0
USA. DOI: https://doi.org/10.1109/Agro-Geoinformatics. Li W., Zhang H., Wang G., Xiong G., Zhao M., Li G., Li R. 2023.
2017.8047016. Deep learning based online metallic surface defect detection
Gaba S., Budhiraja I., Kumar V., Garg S., Kaddoum G., Hassan method for wire and arc additive manufacturing. Robotics
M. M. 2022. A federated calibration scheme for convolu- and Computer-Integrated Manufacturing 80: 102470. DOI:
tional neural networks: models, applications and challeng- https://doi.org/10.1016/j.rcim.2022.102470
es. Computer Communications 192: 144–162. DOI: https:// Liu W., Liu K., Chen D., Zhang Z., Li B., El-Mogy M. M., Tian S.,
doi.org/10.1016/j.comcom.2022.05.035 Chen T. 2022. Solanum lycopersicum, a model plant for the
 Nyarko B.N.E. et al.: Tomato fruit disease detection based on improved single shot detection algorithm 417

studies in developmental biology, stress biology and food seedborne pathogen. Plant Disease 105 (6): 1581–1595.
science. Foods 11 (16): 2402. DOI: https://doi.org/10.3390/ DOI: https://doi.org/10.1094/PDIS-08-20-1732-FE
foods11162402 Picon A., Seitz M., Alvarez-Gila A., Mohnke P., Ortiz-Barre-
Liu B., Luo L., Wang J., Lu Q., Wei H., Zhang Y., Zhu W. 2023a. do A., Echazarra J. 2019. Crop conditional convolutional
An improved lightweight network based on deep learning neural networks for massive multi-crop plant disease clas-
for grape recognition in unstructured environments. Infor- sification over cell phone acquired images taken on real
mation Processing in Agriculture 2023. DOI: https://doi. field conditions. Computers and Electronics in Agricul-
org/10.1016/j.inpa.2023.02.003 (in press) ture 167: 105093. DOI: https://doi.org/10.1016/j.com-
Liu H., Wang D., Xu K., Zhou P., Zhou D. 2023b. Lightweight pag.2019.105093
convolutional neural network for counting densely piled Rashid I., Shoala T. 2020. Nanoactivities of natural nanoma-
steel bars. Automation in Construction 146: 104692. DOI: terials rosmarinic acid, glycyrrhizic acid and glycyrrhizic
https://doi.org/10.1016/j.autcon.2022.104692 acid ammonium salt against tomato phytopathogenic fungi
Ma X., Man Q., Yang X., Dong P., Yang Z., Wu J., Liu C. 2023. Alternaria alternata and Penicillium digitatum. Journal of
Urban feature extraction within a complex urban area with Plant Protection Research 60 (2): 1–11. DOI: https://doi.
an improved 3D-CNN using airborne hyperspectral data. org/10.24425/jppr.2020.133309
Remote Sensing 15 (4): 992. DOI: https://doi.org/10.3390/ Sánchez P., Vélez-del-Burgo A., Suñén E., Martínez J., Postigo I.
rs15040992 2022. Fungal allergen and mold allergy diagnosis: Role
Mohan A., Krishnan R., Arshinder K., Vandore J., Ramanathan and relevance of Alternaria alternata Alt a 1 protein family.
U. 2023. Management of postharvest losses and wastages in Journal of Fungi 8 (3): 277. DOI: https://doi.org/10.3390/
the Indian tomato supply chain – a temperature-controlled jof8030277
storage perspective. Mdpi-Sustainability 15 (2): 1331. DOI: Shi M., He P., Shi Y. 2022. Detecting extratropical cyclones of
https://doi.org/10.3390/su15021331 the northern hemisphere with single shot detector. Mdpi-
Nazari K., Ebadi M. J., Berahmand K. 2022. Diagnosis of alter- Remote Sensing 14 (2): 254. DOI: https://doi.org/10.3390/
naria disease and leafminer pest on tomato leaves using im- rs14020254
age processing techniques. Journal of the Science of Food Sreedevi A., Manike C. 2023. Development of weighted ensem-
and Agriculture 102 (15): 6907–6920. DOI: https://doi. ble transfer learning for tomato leaf disease classification
org/10.1002/jsfa.12052 solving low resolution problems. The Imaging Science Jour-
Nkongho R.N., Ndam L.M., Akoneh N.N., Tongwa Q.M., Nji- nal 71 (2): 1–27. DOI: https://doi.org/10.1080/13682199.20
lar R.M., Agbor D.T., Sama V., Ojongakpa O.T., Ngone A.M. 23.2178605
2023. Vegetative propagation of F1 tomato hybrid (Solanum Thakur P.S., Sheorey T., Ojha A. 2023. VGG-ICNN: A Light-
lycopersicum L.) using different rooting media and stem- weight CNN model for crop disease identification. Multi-
nodal cuttings. Journal of Agriculture and Food Research 11: media Tools and Applications 82 (1): 497–520. DOI: htt-
100470. DOI: https://doi.org/10.1016/j.jafr.2022.100470 ps://doi.org/10.1007/s11042-022-13144-z
Nyarko B.N.E., Bin W., Zhou J., Agordzo G.K., Odoom J., Thakur R., Mohanty S., Sethy P.K., Patro N., Sethy P., Acha-
Koukoyi E. 2022. Comparative analysis of AlexNet, ry A.A. 2023. Detection of tomato leaf ailment using con-
Resnet-50, and Inception-V3 models on masked face rec- volutional neural network technique. p. 193–202. In: Pro-
ognition. p. 337–343. In: Proceedings of “2022 IEEE World ceedings of “Third Mobile and Radio Communications and
AI IoT Congress (AIIoT)”. 6–9 June 2022. DOI: https://doi. 5G Networks. 10–12 June, 2022. Krukshatra, India. DOI:
org/10.1109/AIIoT54504.2022.9817327 https://doi.org/10.1007/978-981-19-7982-8_17
Ouhami M., Hafiane A., Es-Saady Y., El Hajji M., Canals R. Vig S., Arora A., Arya G. 2023. Automated license plate detec-
2021. Computer vision, IoT and data fusion for crop disease tion and recognition using deep learning. p. 419–431 In:
detection using machine learning: a survey and ongoing Proceedings of “Advancements in Interdisciplinary Re-
research. Remote Sensing 13 (13): 2486. DOI: https://doi. search: First International Conference, AIR 2022”. 6–7 May
org/10.3390/rs13132486 2022, Prayagraj, India. DOI: https://doi.org/10.1007/978-3-
Ozbay N., Newman S.E. 2004. Fusarium crown and root rot of 031-23724-9_39
tomato and control methods. Plant Pathology Journal 3 (1): Vishnoi V.K., Kumar K., Kumar B. 2021. Plant disease detection
9–18. DOI: https://doi.org/10.3923/ppj.2004.9.18 using computational intelligence and image processing.
Özbay N., Newman S., Brown W. 2004. Evaluation of Tricho- Journal of Plant Diseases and Protection 128: 19–53. DOI:
derma harzianum strains to control crown and root rot https://doi.org/10.1007/s11119-019-09703-4
of greenhouse fresh market tomatoes. Acta Horticultu- Wei D., Wei X., Tang Q., Jia L., Yin X., Ji Y. 2023. RTLSeg:
rae 635: 635. DOI: https://doi.org/10.17660/ActaHor- A novel multi-component inspection network for railway
tic.2004.635.10 track line based on instance segmentation. Engineering
Peritore-Galve C., Tancos M., Smart C. 2020. Bacterial canker Applications of Artificial Intelligence 119: 105822. DOI:
of tomato: revisiting a global and economically damaging https://doi.org/10.3390/s23041779
 © 2023. This work is licensed under
https://creativecommons.org/licenses/by-sa/4.0/ (the “License”).
Notwithstanding the ProQuest Terms and Conditions, you may use this
content in accordance with the terms of the License.