A Deep Learning Approach to detect the spoiled fruits.

PRIYANKA KANUPURU1, N.V. UMA REDDY2

1Research Scholar, Department of Electronics and Communication Engineering

AMC Engineering College, Affiliated to Visvesvaraya Technological University,

Bangalore – 560083, INDIA

2Professor and Head, Department of AI and ML,

New Horizon College of Engineering,

Bangalore – 560103, INDIA

Abstract: - Fruits are one of the vital sources of nutrients for the mankind and their life span is very less. The

fruit spoilage may occur at various stages such as, at the harvest time, during transportation, during storage etc.

Freshness is a parameter used for accessing the quality of the fruit. About 20% of the harvested fruits are spoiled

due to many factors, before consumption by humans. The spoilage of one fruit has a direct impact on the

neighboring fruits. It is also a one of the indicators that gives an estimation of number of days that a fruit can be

preserved. Early identification of the spoilage helps in taking the appropriate measures for the removal of spoiled

fruits from the whole lot. So that it helps in preventing the spread of spoilage to its adjacent fruits. Deep learning

based technological advancements helps in automatically identifying the spoiled fruits. In this work, internal

quality attributes of the fruit are not taken into consideration for spoilage detection, only the external attributes

are considered. The supervised learning technique is employed for the freshness analysis of two different types

of fruits, Apple and Banana. As the 2 varieties are involved, it is a multiclass classification model with 4 classes.

One shot detection technique is employed to accurately classify among the good fruit and spoiled fruit. Few

images in the dataset are obtained from the kaggle.com and the rest are self - captured images. The dataset is

balanced to avoid the biasness in the model. The model is implemented using Yolov4 and tiny Yolov4 frame

works. These are one shot detection techniques, can be used for real time deployment. The inferences were

obtained on the real time images and video. Confusion matrix is tabulated the performance metrics such as

accuracy, F1 Score and recall are discussed with respect to these two techniques.

Key-Words: - Deep Learning, Fruit Spoilage Detection, Artificial Intelligence, Augmentation, one shot detection

Received: March 27, 2021. Revised: April 15, 2022. Accepted: May 15, 2022. Published: July 8, 2022.

1 Introduction

Fruits are one of the natural sources of food resource

that is derived from plants. Everywhere around the

world fruits are considered to be healthy food. Fruit

consumption by humans gives them a balanced diet.

Due to huge demand, farmers are looking towards

cultivating the fruit crops by employing the internet

of things in agriculture to improve the yield and

productivity [1]. The industries making beverages are

highly dependent on fruits as their major ingredient

[2]. Each variety of fruit has its own lifetime. They

can be degraded quickly if not stored in a proper

manner. With the technological advancements,

application of internet of things in agriculture,

various sensors are used to automate the various

applications in agriculture[3]. The adapted

agricultural practices during the pre-harvest, harvest

and the post-harvest stages also have a great impact

on the productivity of the fruits [4]. Crop

maintenance at each of these stages plays a

predominant role. During the pre-harvest stage the

major influential factors are pesticide or fertilizer

application and weed control. In the harvest stage,

maturity of the fruit and firmness are the key

parameters to be considered. In the post-harvest

stage, storage facilities have a great impact on the life

span of the fruit. Improper maintenance of the crop

under these stages will result in huge loss. Amongst

these factors, pests and diseases are main causes of

low yield [5]. After harvest fruits will continue to

have the respiratory process thereby, they will still

appear fresh for few days. As the days passes fruits

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

will lose its freshness and gradually decay. Apart

from these, poor storage and transportation facilities

contributes to fruit spoilage [6]. Also, there are

chances that the secretions coming out of one spoiled

fruit may damage the entire lot, by spreading through

the neighboring fruits.

At present, Computer vision has its enormous

applications in the field of agriculture, automotive,

education, health care etc. Agri-based computer

vision applications include, identification of plants

and weeds, disease detection in fruit and plant, flower

classification, counting of fruits in a tree for yield,

sensor data analysis related to agricultural parameters

such as temperature, humidity, pH, soil moisture [7].

Machine vision models are the better decision

makers, which has become alternatives to human

beings [8]. In the case of the human being the

decision-making capabilities are dependent on their

naked eye vision. Farmers should have a priori

knowledge of the diseases to classify them based on

their symptoms. Classification can be between the

fruit of same type and the classification between

different species of fruits [9]. The computer vision

techniques extract the features that show the clear

indication of the symptoms to predict the fruit

diseases. These methods rely on the visual features

that appear on the fruit such as shape, color, and

texture [10]. Deep learning framework imitates the

human brain by using the neural networks for data

processing. It requires huge data for training. It gives

an improved performance with different types of

data. The convolution concept was used to determine

the patterns from the image [11]. The deep learning-

based object detection and recognition helps in

building an efficient classification model for

identifying the spoiled fruits from healthy ones.

With this motivation, in this study an automatic

classification of diseased fruits from the healthy ones

was taken into consideration. Banana is a common

man fruit, low cost, multi vitamin rich food [12] and

is available in all seasons. It has rich soluble and

insoluble fiber content, that helps in easy digestion. It

also contains the rich source of nutrients and this fruit

is highly recommended for the people suffering from

anemia. An apple consumption every day, keeps the

doctor away, as per the famous saying, apple and

banana fruits were considered in this paper for

spoilage detection. The traditional transfer learning

approaches classifies only for mutually exclusive

dataset. To classify for the non- mutually exclusive

dataset one shot detection-based frame works are

employed. This paper is organized into six sections.

Second section describes the related work. The model

architecture of the convolutional neural network

(CNN) is explained in third section. The results and

discussions were explained in the fourth section. The

conclusion is given in the fifth section.

2 Related work

Please, Researchers have developed several machine

learning algorithms for object classification using

images. This in turn paved the way for the next

generation artificial neural networks with deep neural

network architectures showing an improved

performance when compared to the traditional state

of art methods.

Mohd Azlan Abu et. al [13] studied the flower

classification using the deep neural network frame on

tensor flow. The dataset consisting of 3670 flower

images were collected from ImageNet website. The

dataset consisting of the five categories comprising

of Daisy, Dandelion, Roses, Sunflower, Tulips were

trained on MobileNet model with an input resolution

of 224x224 RGB images. The trained accuracy

obtained with MobileNet 0.50 was greater than that

of MobileNet 1.00. The model accuracy was 90%.

J.J. Zhuang et.al [14] had developed a machine vision

model to identify the citrus fruits in citrus trees. The

machine vision model consists of illumination

enhancement, foreground region segmentation,

region extraction and recognition. The model was

trained on the self captured images. The bounding

boxes with the minimum enclosure are drawn over

the detected fruits in the test data. The model was

trained with 100 images. The F1 score obtained was

0.91. David Ireri et. al[15] had proposed a computer

vision model to detect the defects in tomatoes and

also to perform tomato grading. 200 tomatoes of

varying composition of defects were chosen and they

were transformed into L, A, B colour space. From

each of these mean, range and standard deviation was

considered to analyse the colour features, further

shape features and text features were analysed.

Support vector machine based were used for

recognition. For the testing purpose the images are

captured with the hikivision camera, connected to

computer with Ethernet cable. The model acuuracy

was 0.95, root mean square error was 33.5.

Kyamelia Roy et. al [16] developed a frame work

to classify the healthy apple and rotten apple. The

dataset was obtained from kaggle.com. data pre-

processing was performed to generate the binary

masks. The enhance UNet was built using the U-Net

frame to improve the accuracy. The model was

trained with GPU Tesla K80 in google colab. The

UNet model achieved 95.36 % validation accuracy,

where with the enhanced UNet model the validation

accuracy is 97.54. Shiv Ram Dubey et. al [17]

developed an apple disease classification model to

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

predict the Apple Blotch, Apple scab, , Apple rot and

Normal Apple. In the initial stage, extraction of the

region of interest was performed, followed by feature

map generation. In the next step distinctive features

were extracted. Multi class support vector machine

was used for classification. The model achieved an

accuracy of 95.94%. Yunong Tian et. al [18]

developed a multi class classification model to

categorize between the defected and the healthy

leaves and fruit. Out of the 11 classes, three of them

belong to healthy leaves and fruit categories and the

rest of them belong to the diseased leaves and fruit

categories. Cycle-GAN was used for augmentation of

images. The dataset is trained over Multi-scale Dense

Inception-V4, obtained accuracy wad 94.31 % and

Multi-scale Dense Inception-ResNet-V2 algorithms

with an accuracy of 94.74%. Poonam Dhiman et. al

[19], identified the disease severity level in citrus

fruits at four levels being low, medium, high and

healthy by applying the transfer learning on the

VGGnet, giving an accuracy of 99% with severity

level low, 98% with high severity level, 97% with

medium and 96% with healthy fruit. Benjamin

Doh,et.al [20] used svm and ANN to identify the

citrus fruit disease based on its characteristics such as

color, holes, texture and morphology. The average

accuracy of ANN was 89% and 87% with SVM.

Inkyu Sa et. al [21], developed a model with

transfer learning on Faster Region-based CNN to

detect the presence of fruits in a farm. The

multimodal information comprising of Near Infra red

and RGB, with 4 channels was given as an input to

the network. The bounding box annotation was

performed on the images. The model was trained on

NVIDIA GPU. The F1 score obtained was 0.83.

Heena Shaikh et. al [23], used a six class classifier to

classify among apple, pear and bear healthy ones and

diseased fruits. In the first step images were

annotated using LabelImg in pascal voc format to

generate an XML file. 344 images were trained on

the faster R- CNN model. The average accuracy was

85. Jamil Ahmad et. al [20], plum images was

captured using mobiles phones are fed to the fine

tuned CNN to train the model for predicting 5 classes,

one class belong to healthy and the other classes

belong the defected classes. The overall performance

of the model is 88.42%. Ganeshan Mudaliar et. al

[24] developed a model to study the classification of

ripen tomato and rotten tomato. The data set was

taken from kaggle and trained over 500 images with

the mobile net. The image output size from pre-

processing stage is 32x32, which is then fed to the

mobile net. The accuracy of this model is 98.74.

Guichao Lin et. al [25], presented a shape matching

model based on sub fragment detection and

aggregation. Aggregation was done by eliminating

the false positives through SVM classifier. The

experimental results achieved precision in the range

0.783 – 0.919 for 6 classes.

From the related work it is understood that as the

size of the network increases, the training time

increases, but the model will produce an improved

accuracy when compared to the lighter model. Deep

learning models requires enormous data when

compared to computer vision techniques, but greater

performance can be achieved. One shot YOLO object

detection has CNN as backbone, used for real time

deployment for object detection.

3 Model Architecture

Convolutional neural network, being a deep learning

algorithm in which, based on the objects in the input

image, the learning weights are modified. With the

help of the filters, CNN can easily obtain the spatial

and temporal dependencies.

Convolution, ReLu, Pooling and fully connected

layers forms the building blocks of CNN. During this

process the dominant features are extracted. Noise

suppression is done in the max pooling layer. In the

fully connected layer, the nonlinear combination of

the dominant features are represented at the output.

The basic CNN architecture block diagrams are

shown in the following figure 1[8]. LeNet, AlexNet,

VGG Net, ResNet, architectures were built using the

basic blocks of CNN.

In the convolution layer output is given

mathematically as the formulation is defined by (1).

where k is the CONV layer, the output feature vector

for layer k was denoted by Xnew.

 



  (1)

 and  represents the elements of the filter and

bias respectively, denotes the activation function.

Rectified Linear unit activation function was applied

to increase the nonlinear properties of the network.

During this stage, the negative input values are

replaced by zero. In the pooling layer, based on the

chosen stride value and the window size the max

pooling or the average pooling is performed. The

feature map obtained from this form the input to the

subsequent layers. The above steps are repeated for

all the layers.

The flattened input is given to the fully connected

layer, to which softmax activation is applied to

produce the output in terms of probabilistic value for

the each of the classes. This model is implemented in

python with tensorflow and keras libraries.

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Fig.1: Basic Convolutional Neural Network Architecture

Fig. 2 : Block diagram of an object detector.

3.1 Object detector

In general an object detector consists of the following

input, backbone neck and head. The objector detector

takes the input as image, patch, image pyramid. Back

bone generally contains an architecture built on CNN

model. The neck contains the path – aggregation

blocks. It forms the feature pyramid network from the

backbone Head contains dense prediction of one

stage and sparse prediction consisting of two stages.

Generally used detectors for one stage are single shot

detector (SSD) and YOLO. Two stage detector model

uses faster RCNN based models. The object detector

diagram is shown in the figure 2 above [26].

3.2 Faster RCNN Architecture:

The faster RCNN consists of, Fast RCNN detector

with VGG as the backbone to obtain the object

features, a region proposal network to mark the

bounding boxes indicating the possible objects in an

image. The classification layer is responsible for the

class predictions. The width and height of the

bounding box can be obtained from the regression

layer. The size of the anchor boxes may vary from

128, 512,1024, the aspect ratio 2:1, 1:1, 1:2. It uses

multitask loss function [27]. The overview of faster

RCNN is shown in figure 3 below [27].

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Fig.3: Block diagram of Faster R-CNN unified Network.

Fig. 4: YOLOV4 block diagram.

3.3 YOLOV4 Architecture

YOLO V4 is a simplified network with a reduced

number of learnable parameters when compared the

traditional CNN models. You Look Only Once, as

indicated by the name, it is a one-shot detector. It is

built on the cross stage partial architecture, which is

built on the Dense Net, called as CSPDarknet53[28].

It not only does image classification but also

performs the object localization. It is represented by

a vector consisting of probability of class, centre (x,y)

coordinates of the bounding box, height, width of the

bounding box, probability of the bounding box

classes. The YOLOV4block diagram is shown in the

figure 4 above [26].

In the neck part, concatenation was used for the

Path Aggregation Network (PAN). Max pooling was

performed in the spatial pyramid pooling. Further bag

of freebies (BoF) methods helps to improve the

accuracy without putting an additional overhead of

interference. Bag of Specials (BoS) consists of Mish

activation, Cross- stage partial connections and

weighted residual connections. Mish activation

function will be similar to ReLu and Swish. During

the training phase the weights are modified

depending on the Complete Intersection Over Union

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

(CIoU). The loss(L) is mathematically given in

equations (2,3) as [25-26]

L = S(ß, ßgt) + D(ß, ßgt) + V(ß,ßgt) (2)

IoU = Intersction Area / Union Area (3)

Where ß is the bounding box, S indicates the

overlap area, ßgt is the ground truth bounding box.

Analysing the CIoU loss helps in maintaining the

consistency of the bounding box aspect ratio. The

non-maximum suppression filters the additional

bounding boxes on the same object and the keeps

only a single bounding box. The uses the IoU and the

distance between the centers of two consecutive

boxes to determines the redundant enclosures.

3.4 YOLOV4 – Tiny Architecture:

YOLOV4 tiny architecture is built from YOLOV4.

This tiny model detects the objects at a faster rate

when compared to YOLOV4. Since it is a lightweight

model, it will be easily deployed on the real time

applications using embedded systems [29-31]. In the

backbone it uses CSPDarknet53-tiny. In the cross

stage partial network, it employs CSP block, in the

cross stage residual edge which divides the feature

map into two parts. So that it increases the correlation

difference in the gradient information. To reduce the

computational overhead it uses Leaky ReLu

activation function. Mathematically the Leaky ReLu

activation is given by [29]

󰇫



 (4)

Where 󰇛 󰇜

Table 1: YOLOV4 tiny network structure.

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Fig.5: Block diagram of YOLOV4 tiny model.

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

The feature pyramid network with different scales

are used to increase the object detection speed. In the

tiny model SPP and PAN are not used. The detections

are predicted using 13x13 and 26x 26 feature maps.

The block diagram of YOLOV4-tny model is shown

in the figure 5 below [32].

Initially the image is divided into SxS grids, in

each grid the objects are detected using bounding

boxes. Depending on the availability of the objects in

the grid, SxSxB bounding boxes are generated. This

process is repeated for the entire image. The object

is predicted only if the centre of the object related

bounding box is present inside the grid. The

bounding box is retained only if it satisfies the

confidence threshold value. Then intersection over

unions, non-maximum suppression is applied same

as YOLOV4 to retain a single bounding box over a

single image. The network structure of YOLOV4 tiny

model is shown in Table.1 above [33].

The average precision in YOLOV4 is increased by

10% when compared to YOLOV3. The YOLOV4

contains 139 convolutional layers whereas YOLOV4

tiny contains 29 convolutional layers. The accuracy

of the YOLOV4 tiny model is two-third of YOLOV4

model on MS COCO dataset. YOLOV4 tiny is

preferable in applications where there is a

requirement of faster processing in real in real time.

4 Results and Discussion

Two types of fruits, apple and banana, which are

healthy and diseased were considered for the study.

The classification and localization were performed

using YOLOV4 tiny and YOLO algorithms. This is

a multi-class classification problem with four classes

being good apple, bad apple, good banana and bad

banana. A total data set of 800 images were used. The

complete data set was split into train and test. A total

of 800 images with 200 images from each class are

collected. Few images were captured manually using

a mobile phone and few are collected from

kaggle.com /fruit 360. The same data set trained over

two frame works YOLOV4 tiny and YOLO. The

dataset was annotated with the bounding boxes using

lablelImg Software. The labelImg generates the .txt

file consisting of data related to class and bounding

box coordinates. The data augmentation was

performed on the images to increase the data set. The

images are rotated by 90 degrees and 270 degrees.

The following figures [6 -7] gives an overview of the

loss in two models. The model was trained using

google colab pro. At the end of 6000 iterations the

current average loss of the two models is shown in

the Table 2. below:

Since YOLOV4 tiny is a lightweight model in

which the loss convergence is faster. Since the

number of convolution layers in the tiny yolov4 less,

the trainable parameters is also less. Therefore, the

time taken by for each iteration is less when

compared to YOLOV4 model.

Fig. 6 YOLOV4 loss convergence with number of iterations during training

Table 2: Average loss during training.

Model

Current average loss at 6000th iteration

Time taken for 6000 iterations

YOLOV4

0.3101

12 hours

YOLOV4 tiny

0.0530

2 hours

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Fig. 7 YOLOV4 tiny model loss convergence with number of iterations during training.

Table 3: Predictions on test images

Sno

Test image

Predicted class

YOLOV4

YOLOV4tiny

Good apple

Bad apple

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Bad apple

Bad banana

Good banana

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

Fig.8(a, b, c, d): Extracted frames of video prediction on YOLOV4 tiny.

Fig.9 (a, b, c, d): Extracted frames of video prediction on YOLOV4.

Table.3 shows image predictions, figures 8, 9 shown

above describes the predictions on the video, few

frames are extracted for both the YOLOV4 tiny and

YOLOV4 models. Green colored bounding box

indicates the good banana, the blue colored bounding

box indicates the bad apple, pink color bounding box

indicates that the predicted object is a good apple.

The confusion matrix is tabulated to describe the

performance of the model for the test images. It gives

the analysis about the predicted data against the target

data. The Table 4 and Table 5describes the confusion

matrix for fruit classification model with YOLOV4

and YOLOV4 tiny algorithms from which the

accuracy and F1 score was calculated.

Table 4: Confusion matrix of the fruit classification model with YOLOV4 algorithm

Predicted Outputs

Actual Class

Good

apple

Bad

apple

Good

banana

Bad

banana

Good apple

Bad apple

Good banana

Bad banana

Table 5: Confusion matrix of the fruit classification model with YOLOV4tiny algorithm

Predicted Outputs

Actual Class

Good

apple

Bad

apple

Good

banana

Bad

banana

Good apple

Bad apple

Good banana

Bad banana

The results obtained from Table 4 and Table5 shows

that there is no misclassification between the apple

and banana with both the models. The distinctive

features between the apple and banana are more

therefore the predictions does not classify the apple

as banana and vice-versa. In the case of classification

between the same species being healthy and spoiled

there is more correlation. Since the tiny YOLOv4

model has the less architecture when compared to the

full model, the misclassification is more between the

same species. From the confusion matrix the

accuracy, F1 score, Recall are calculated as shown in

the equations 5,6,7,8.

(c)

(d)

(c)

(d)

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

(5)

(6)

(7)

(8)

From Table 6, it is observed that the as the network

architecture is more, the number of trainable

parameters increases, performance metrics

parameters shows improvement. The accuracy and

F1 score of the YOLOV4 algorithm is higher in

comparison with YOLOV4 tiny when trained for

6000 iterations.

Table 6: Comparison of performance metrics for YOLOV4 and YOLOV4tiny.

Accuracy

Precision

Recall

F1score

YOLOV4

0.94

0.938

0.92

0.927

YOLOV4tiny

0.87

0.86

5 Conclusion

In this paper apple and banana are classified between

the healthy and the diseased ones. Deep learning

YOLO algorithms will give higher accuracy and

precise object detection in comparison with

traditional CNN models. The dataset is trained on

YOLOV4 and the light weight model YOLOV4 tiny.

Since YOLOV4 tiny has a smaller number of

convolution layers in comparison with the YOLOV4,

it consumes less time for training. For this dataset,

after 6000 iterations the average loss in YOLOV4

tiny is less when compared to YOLOV4. Predictions

are done for the same set of images in YOLOV4 and

YOLOV4 tiny, it was observed that accuracy of the

YOLOV4 tiny model is slightly lesser than the

YOLOV4. For a given test image the confidence

score of identifying the object in YOLOV4 is higher

than YOLOV4 tiny, but faster predictions occur in

YOLOV4tiny as it a lightweight model so it can be

easily deployed in embedded systems for real time

applications. Therefore, there is a trade- off between

the time consumption, confidence level in

classification and accuracy. For a small-scale

applications YOLOv4 tiny is a good choice. In

Future, the model can be extended for other category

of fruits by re-training the model with relevant

dataset and can be deployed on embedded platform

for real time fruit grading.

Declaration of Interests

The authors declare that there is no conflict of

interest.

References:

[1]. Priyanka Kanupuru, N. V. Uma Reddy, "Survey

on IoT and its Applications in Agriculture," 2018

International Conference on Networking,

Embedded and Wireless Systems (ICNEWS), pp.

1-5, 2018. doi: 10.1109/ICNEWS.2018.8903969.

[2]. L. Wu, H. Zhang, R. Chen and J. Yi, "Fruit

Classification using Convolutional Neural

Network via Adjust Parameter and Data

Enhancement," 2020 12th International

Conference on Advanced Computational

Intelligence (ICACI), pp. 294-301, 2020. doi:

10.1109/ICACI49185.2020.9177518.

[3]. Kanupuru Priyanka, Nadig Vijayendra Reddy

Uma Reddy, “Multi node based smart monitoring

system with motor dry run avoidance for

sustainable agriculture”. Instrumentation Mesure

Métrologie, Vol. 20, No. 1, pp. 39-47, 2021.

https://doi.org/10.18280/i2m.200106.

[4]. Vishal Meshrama , Kailas Patil , Vidula

Meshrama , Dinesh Hanchate b , S.D. Ramkteke,

“Machine learning in agriculture domain: A state-

of-art survey”, Artificial Intelligence in the Life

Sciences, Elsevier, Vol. 01, pp. 25761-25783,

2021.

https://doi.org/10.1016/j.ailsci.2021.100010.

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

[5]. Guy Blomme, Miguel Dita, Kim Sarah Jacobsen,

Luis Pérez Vicente, Agustin Molina, Walter

Ocimati, Stephane Poussier, Philippe Prior,

(2017) “Bacterial Diseases of Bananas and

Enset: Current State of Knowledge and

Integrated Approaches Toward Sustainable

Management”, Frontiers in Plant Science, Vol.

08, pp. 1 – 25.

https://doi.org/10.3389/fpls.2017.01290.

[6]. Arif U Alam , Pranali Rathi, Heba Beshai ,

Gursimran K,Sarabha and M, Jamal Deen, “Fruit

Quality Monitoring with Smart Packaging”,

Sensors, vol.21, pp. 1 – 20, 2020,

https://doi.org/10.3390/s21041509.

[7]. Juliana Aparecida Fracarolli, Fernanda

Fernandes Adimari Pavarin , Wilson Castro and

Jose Blasco, “Computer vision applied to food

and agricultural products”, Revista Ciência

Agronômica, , Special Agriculture 4.0, vol.51,

pp. 1 – 20, 2020, doi:10.5935/1806-

6690.20200087.

[8]. Dayanand Savakar, “Identification and

Classification of Bulk Fruits Images using

Artificial Neural Networks”, International

Journal of Engineering and Innovative

Technology (IJEIT), vol.01,No. 3, pp. 36-40,

2012.

[9]. Rojas-Aranda J.L., Nunez-Varela J.I., Cuevas-

Tello J.C., Rangel-Ramirez G. Fruit

Classification for Retail Stores Using Deep

Learning. In: Figueroa Mora K., Anzurez Marín

J., Cerda J., Carrasco-Ochoa J., Martínez-

Trinidad J., Olvera-López J. (eds) Pattern

Recognition. MCPR 2020. Lecture Notes in

Computer Science, vol 12088, 2020. Springer,

Cham., https://doi.org/10.1007/978-3-030-

49076-8_1.

[10]. Muhammad Attique khan, Tallha Akram,

Muhammad Sharif, Tanzila Saba, “Fruits

diseases classification: exploiting a hierarchical

framework for deep features fusion and

selection”, Multimedia Tools and Applications,

Springer, Vol. 79, pp. 25763–25783, 2020.

https://doi.org/10.1007/s11042-020-09244-3 .

[11]. Sai Sudha Sonali Palakodati , Venkata

RamiReddy Chirra, Yakobu Dasari , Suneetha

BullaYk, “Fresh and Rotten Fruits Classification

Using CNN and Transfer Learning”, Revue

d'Intelligence Artificielle, vol.34, No.05, pp.

617-622, 2020.

https://doi.org/10.18280/ria.340512.

[12]. K. Lakshmi Narayanan , R. Santhana Krishnan

, Y. Harold Robinson , E. Golden Julie , S. Vimal

, V. Saravanan , and M. Kaliappan, Banana Plant

Disease Classification Using Hybrid

Convolutional Neural Network", Computational

Intelligence and Neuroscience , vol. 2022,

Article ID 9153699, 13 pages, 2022.

https://doi.org/10.1155/2022/9153699.

[13]. Mohd Azlan Abu , Nurul Hazirah Indra , Abdul

Halim Abd Rahman , Nor Amalia Sapiee and

Izanoordina Ahmad, “A study on Image

Classification based on Deep Learning and

Tensorflow”, International Journal of

Engineering Research and Technology, vol.12,

No.4, pp. 563-569, 2019.

[14]. J.J. Zhuanga , S.M. Luoa , C.J. Houa , Y. Tanga

, Y. Heb , X.Y. Xuec, “Detection of orchard

citrus fruits using a monocular machine vision-

based method for automatic fruit picking

applications”, Computers and Electronics in

Agriculture, vol.152, pp.64-73, 2018.

https://doi.org/10.1016/j.compag.2018.07.004.

[15]. David Ireri, Eisa Belal, Cedric Okinda, Nelson

Makange, Changying Ji Gzjf, “A computer

vision system for defect discrimination and

grading in tomatoes using machine learning and

image processing”, Artificial Intelligence in

Agriculture, vol 2, pp 28 – 37, 2019.

https://doi.org/10.1016/j.aiia.2019.06.001.

[16]. Roy, K., Chaudhuri, S.S. & Pramanik, S. “Deep

learning based real-time Industrial framework for

rotten and fresh fruit detection using semantic

segmentation” Microsystem Technologies, Vol.

27, pp. 3365–3375, 2021.

https://doi.org/10.1007/s00542-020-05123-x

[17]. Shiv Ram Dubey, Anand Singh Jalal, “Apple

disease classification using color, texture and

shape features from images”, Signal, Image and

Video Processing volume, Springer, Vol. 10, pp.

819–826, 2016. https://doi.org/10.1007/s11760-

015-0821-1.

[18]. Tian Yunong, Li En, Liang Zize, Tan Min, He

Xiongkui, “Diagnosis of Typical Apple

Diseases: A Deep Learning Method Based on

Multi-Scale Dense Classification Network”,

Frontiers in Plant Science, Vol. 12, , pp. 1-12,

2021.

https://doi.org/10.3389/fpls.2021.698474.

[19]. Dhiman, Poonam, Vinay Kukreja, Poongodi

Manoharan, Amandeep Kaur, M. M.

Kamruzzaman, Imed B. Dhaou, and Celestine

Iwendi. "A Novel Deep Learning Model for

Detection of Severity Level of the Disease in

Citrus Fruits" Electronics, Vol 11, no. 3: 495, pp

1-14. 2022.

https://doi.org/10.3390/electronics11030495.

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022

[20]. B. Doh, D. Zhang, Y. Shen, F. Hussain, R. F.

Doh and K. Ayepah, "Automatic Citrus Fruit

Disease Detection by Phenotyping Using

Machine Learning," 2019 25th International

Conference on Automation and Computing

(ICAC), pp. 1-5,2019. doi:

10.23919/IConAC.2019.8895102.

[21]. Inkyu Sa, Zongyuan Ge, Feras Dayoub, Ben

Upcroft, Tristan Perez, Chris McCool,

“DeepFruits: A Fruit Detection System Using

Deep Neural Networks” sensors, vol.16, No. 08,

pp. 1-23, 2016.

https://doi.org/10.3390/s16081222.

[22]. Heena Shaikh, Yashashree Wagh, Siddhesh

Shinde, Shankar M. Patil, “Classification of

Affected Fruits using Machine Learning”,

International Journal of Engineering Research &

Technology (IJERT), vol.9, no.7, pp. 1541-

1547,2021.

[23]. Jamil Ahmad , Bilal Jan , Haleem Farman ,

Wakeel Ahmad and Atta Ullah, “Disease

Detection in Plum Using Convolutional Neural

Network under True Field Conditions”, Sensors,

Vol. 20, no.9, pp. 1–18, 2020.

https://doi.org/10.3390/s20195569

[24]. Ganeshan Mudaliar, Rashmi Priyadarshini B.K,

“A Machine Learning Approach for Predicting

Fruit Freshness Classification”, International

Research Journal of Engineering and Technology

(IRJET), Vol. 08, no. 05 , pp. 2979-2983,2021.

[25]. Guichao Lin, Yunchao Tang, Xiangjun Zou,

Jiabing Cheng, Juntao Xiong, “Fruit detection in

natural environment using partial shape matching

and probabilistic Hough transform”, Precision

Agriculture, Vol.21, pp. 160–177, 2020.

https://doi.org/10.1007/s11119-019-09662-w.

[26]. Alexey Bochkovskiy, Chien-Yao Wang, Hong-

Yuan Mark Liao, “YOLOv4: Optimal Speed and

Accuracy of Object Detection”,

arXiv:2004.10934v1,2020.

[27]. Shaoqing Ren, Kaiming He, Ross Girshick, Jian

Sun, Faster R-CNN: Towards Real-Time Object

Detection with Region Proposal Networks,

arXiv:1506.01497v3, pp 1-14,2016.

[28]. Protik, Adban & Hossain, Amzad & Siddique,

Shahnewaz,“Real-time Personal. Protective

Equipment (PPE) Detection Using YOLOv4 and

TensorFlow”, 2021 IEEE Region 10 Symposium

(TENSYMP), pp. 1-6, 2021. DOI:

10.1109/TENSYMP52854.2021.9550808.

[29]. Anna Kuznetsova, Tatiana Maleva and

Vladimir Soloviev, Using YOLOv3 Algorithm

with Pre- and Post-Processing for Apple

Detection in Fruit-Harvesting Robot”,

Agronomy, Vol. 10, no. 1016, pp. 1-19, 2020.

doi:10.3390/agronomy10071016.

[30]. WenYu Feng,1 YuanFan Zhu,2 JunTai Zheng,1

and Han Wang, “Embedded YOLO: A Real-

Time Object Detector for Small Intelligent

Trajectory Cars “Mathematical Problems in

Engineering, vol. 2021, Article, pp. 1-11, 2021.

https://doi.org/10.1155/2021/6555513.

[31]. Dongwei Guo, Lufei Cheng, Meng

Zhang,Yingying Sun, “Garbage detection and

classification based on improved YOLOV4”,

Journal of Physics: Conference Series, 2024

012023, pp. 1-8, 2021. doi:10.1088/1742-

6596/2024/1/012023.

[32]. Zicong Jiang1 , Liquan Zhao , Shuaiyang Li

And Yanfei Jia, “Real-Time Object Detection

Method based on improved YOLOv4-tiny”,

Computer vision and pattern recognition, pp.1-

11, 2020.

https://doi.org/10.48550/arXiv.2011.04244 .

[33]. Silva, L.A.; Sanchez San Blas, H.; Peral García,

D.; Sales Mendes, A.; Villarubia González, G.

“An Architectural Multi-Agent System for a

Pavement Monitoring System with Pothole

Recognition in UAV Images”, Sensors vol. 20,

no. 21, pp. 1-23. 2020,

https://doi.org/10.3390/s20216205.

Contribution of individual authors to

the creation of a scientific article

(ghostwriting policy)

Priyanka Kanupuru contributed for the execution of

algorithms and analysis of results.

N V Uma Reddy contributed for the analysis of

results.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the Creative

Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en_US

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.10

Priyanka Kanupuru, N. V. Uma Reddy

E-ISSN: 2415-1521

Volume 10, 2022