Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Tobacco Disease Detection and Classification for

Grading System Using Convolutional Neural Network
Pempho Jimu Dr. G. Glorindal Selvam
Student: Department of Computer Science Project Supervisor: Department of Computer Science
DMI-St. Eugene University, Zambia DMI-St. John the Baptist University, DMI-SJBU
Lilongwe, Malawi Lilongwe, Malawi

Abstract:- Tobacco being one of the perennial crops that application to another. Usually this is the case because a
are the major source of forex in Malawi. But due to the digital image processing system is hard coded to only be
decrease in agricultural extension workers specializing in effective and efficient in only one area of application and
Tobacco this paper proposes a remedy for tobacco disease might not perform well in another field. This problem is
detection and grading using a convolution neural network corrected by adding highly advance digital image processing
(CNN). This paper proposes the use of CNN as a machine system with artificial intelligence, specifically neural
learning algorithm for the training a system to detect networks due to their capability of learning new tasks by
tobacco diseases while grading the tobacco. This paper specifying urgings in the algorithm. Digital image processing
also compares CNN algorithm with other neural network has a lot of algorithms which use different techniques but
algorithms in terms of sensitivity, specificity and other share common steps to achieve a given objective whether it is
parameters such that the CNN proves to be efficient and image compression or image recognition. The fundamental
effective hence offering higher performance of the steps in digital image processing image acquisition, image
algorithms. This research work will be designing and enhancement, image restoration, color image processing,
developing a learning model, which will classify, grade wavelets and multiresolution processing, compression,
and detect the disease in the tobacco plant leaf by using morphological processing, segmentation, representation and
the convolutional neural network based on the type of the description, and object recognition. This paper involves the
tobacco Leaf. As these convolutional networks work with usage of digital image processing together with neural
a large number of images for the training of the model, networks in the recognition and identification of tobacco
real-time images will be collected from the reliable source diseases [15].
such as Agricultural Research and Extension Trust
(ARET) and Tobacco Auction Flow Point respectively. II. RELATED WORKS

Keywords:- Image Processing, Convolutional Neural Agricultural equipment is evolving specifically precise
Network (CNN), Leaf Detection, Agricultural Research and prediction of diseases in plants [1]. This has facilitated very
Extension Trust (ARET), Tobacco Classification, Tobacco much in the objective of attaining food security in a lot of
Grading countries and making sure that farmers have the maximum
quality of produce from the farm. The capability to detect
I. INTRODUCTION diseases and to grade tobacco reduces subjective interference
of human beings who are fallible. Expert systems were one of
The effective preservation of plant diseases is closely the necessary steps toward automation of detection of tobacco
related to bearable agriculture and climate variation [8]. diseases and grading, but they were not as accurate and fast
Digital image processing has been applied in a variety of as those of neural networks [2]. Expert systems are rule-based
fields especially in medicine and agriculture. Digital image where by the algorithm is based on conditions defined by a
processing is used in a numerous-fields and it helps in programmer or system engineer. A field as complex as
automation of various processes requiring human being visual medicine it is almost impossible for a system developer to
perception which is effectively side stepped by the inception anticipate all possible conditions without missing out. This
of digital image processing. Digital image processing is very gives enough of a reason why machine learning using neural
important because it involves the use of machines which are networks fits better for problems such as detection and
accurate every time they are running since they do not get prediction of diseases. In the paper [3] convoluted neural
tired despite the long time spent running. Another importance networks (CNN) are used as the basis for training a system.
of digital image processing is the capability of using a high Deep neural networks (DNN) are the most basic but they have
computation power obtained from advanced computers and a disadvantage with vanishing gradients. In order to overcome
computing concepts such quantum computing which can help the drawbacks of such traditional method, new techniques are
in performing highly complex tasks within a small amount of needed to predict the grading process. This has been proven
time. Not only that, but also, application in a wide variety of by the previous studies [4,9], where they applied image
areas which makes it very flexible. Although digital image processing and machine learning techniques to develop an
processing acquisition is much of attention in technological automatic grading system for flue-cured tobacco leaves. They
developments, it is also very incomplete especially when it projected algorithms to extract the leaves’ color, shape,
comes to interoperability of a given system from one area of quality, injury and waste tolerance, stalk point, and maturity.

IJISRT22SEP836 www.ijisrt.com 1120

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Relative to classification, machine learning algorithms like Below is a diagram that shows a neural network in
Fuzzy [4], Support Vector Machine [7], K-Nearest Neighbor forward propagation with three layers:
[6], Generalized Regression Neural Network [9], and
Backpropagation Neural Network [8] were used to classify
tobacco leaves. Because the classification features were
selected based on human experience, these approaches
enhanced the accuracy of tobacco grading. However, the
sorting accuracy is still not high enough and is vulnerable to
artificial feature selection. Developed in recent years,
Convolutional Neural Networks (CNNs) have become
popular especially in computer vision because they achieved
remarkable performances on different tasks such as image
classifications in the field of agriculture [9,11]. CNN is an
end-to-end pipeline that can automatically discover the useful
features for classification, whose advantage lie in local 3-D
coherence in the input images, which allow them to have
fewer bulks as some constraints are shared. Inspired by the
development of CNN in image-based acknowledgment and
not much studies on the application of CNN in tobacco leaf
grading, this study attempts to use CNN in grading and
classification of tobacco leaves and compare with the existing
models [10]. This paper endorses convolution neural
networks (CNN) which do not jump certain layers without
suppressing the reduction endangered gradients which
Fig 1:- neural network with three layers
destructively impact the overall training of neural networks as
they are being used. CNN is also more significant because the
 Image Acquisition
training is relatively faster than the DNN and is more novel in
Digital images are acquired manually by users using
finding a lot of connections within the neural network.
either the user interface during the process of detecting
tobacco disease and grading tobacco within an image and
III. METHODOLOGY
images are also acquired in a bulk in the form of a dataset as
required by the code in MATLAB by a specified path.
Before you begin to format your paper, first write and
Acquisition conveys images to the image enhancement.
save the FORWARD PROPAGATION:
 Image Enhencement
A neural network given the weight of W^ (l-1, l) for
Image enhancement converts the scale of an image or
connection from layer l-1 to layer l and the weight W^(l-21, l)
images to a uniform scale so as to minimize errors during
for connection from l-2 to l. This makes the forward
classification and to avoid feeding a lot of errors to a given
propagation to be as shown below:
neural network [7]. The process is not complicated because the
acceptable scale of both horizontal and vertical axis is the
𝒂𝒍 ≔ 𝒈(𝑾𝒍−𝟏,𝒍 . 𝒂𝒍−𝟏 + 𝒃𝒍 + 𝑾𝒍−𝟐,𝒍 . 𝒂𝒍−𝟐 ) defined initially in the code.
𝒈(𝒁𝒍 + 𝑾𝒍−𝟐,𝒍 . 𝒂𝒍−𝟐 )
 Feature Extraction
Whereby the variables are: In as much as accelerative propagation is concerned,
a^l the activations of neurons in a layer l feature extraction is solely a result of the neural network’s
g the activation function of layer l forward propagation. The system uses activation of neurons
𝒁𝒍 = 𝑾𝒍−𝟏,𝒍 . 𝒂𝒍−𝟏 + 𝒃𝒍 . whereby each and every layer has its own feature to extract.
Since this system uses RESNET-50 which has 50 layers and
This is the forward propagation of the system of a the algorithm does not jump other layers making it to be
convolution neural network. A general case of layers more functionally a convolution neural network [4,5].
than three, the equation for the neuron activation is represented
as follows:  Classification
𝒌 Classification is nothing more than the assigning of a
meaningful label to a given image by mapping a selected
𝒂𝒍 ≔ 𝒈(∑ 𝑾𝒍−𝒌,𝒍 . 𝒂𝒍−𝒌 ) image feature weighted probabilities and comparing them to
𝒌−𝟐 each of the feature weighted probabilities by the datasets. If
there is congruency within the values a corresponding label is
Whereby k is the number of layers within a given neural assigned to the selected image during classification.
network.

IJISRT22SEP836 www.ijisrt.com 1121

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Below is a diagram that shows how the modules explained above are interconnected together.

Fig 2:- module description figure

Fig 3:- tobacco disease prediction flow diagram

IJISRT22SEP836 www.ijisrt.com 1122

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 4:- tobacco disease prediction implementation

IV. RESULTS AND DISCUSSION Accuracy is given by the formula:

After the text edit has been completed, the paper is ready 𝑇𝑁+𝑇𝑃
Accuracy=
Using MATLAB quantitative results were collected with the 𝑇𝑁+𝑇𝑃+𝐹𝑃+𝐹𝑁
help of high-end performance computers so as to have a
detailed account of how the residual neural network algorithm Whereby, TN means true negative
works and performs as compared to other types of neural TP means true positive,
network algorithms. FP means false positive,
and FN means false negative.
Figure 5 shows that CNN is the best algorithm in terms
of accuracy as compared to both deep neural network (DNN)
due to the missing or disappearing gradients because in the
DNN measurements are taken frequently hence little to no
change is observed.

IJISRT22SEP836 www.ijisrt.com 1123

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Figure 8 portrays the F-measure between DNN, RNN
accuracy between algorithms and CNN. From the graph it shows that both DNN and RNN
have a similar F-measure while the CNN has a different F-
92 measure which is of a higher value than both RNN and DNN.
91.8
Accuracy

91.6
F-measure between algorithms
91.4
71
91.2
70
70

F-measure
91
DNN RNN CNN
69
Algorithm
68
Fig 5:- accuracy between algorithms 67
DNN RNN CNN
From figure 6 sensitivity of the CNN algorithm is
exponentially greater than that of both RNN and DNN Algorithm
because the CNN does not jump layers like the RNN and it
takes into consideration small difference between objects or Fig 8:- F-measure between algorithms
features.
V. CONCLUSION

sensitivity between algorithms This paper uses CNN with the help of RNN in such a
way that the traditional structure and partition memory works
85 as feedback loops. RNN (residual neural network) was
designed and developed using MATLAB Resnet-50 and it
80 successfully predicted tobacco diseases upon training it with a
Sensitivity

given dataset. The RNN algorithm is proven to be better than

75 the DNN algorithm and it is a little less than the CNN
algorithm. The findings show that RNN is better than current
70 deep and machine-learning models. The results for the
detecting and grading of leaf samples are also more accurate
65 than other approaches. The use of CNN reveals that segmented
DNN RNN CNN and non-segmented experimental results of images. The
Algorithm classification of images can effectively be as high as 97.23%
for grading tobacco and it took 6.78ms to classify a single leaf
Fig 6:- sensitivity between algorithms for a tobacco. This method and approach makes easier for
grading and detecting the disease effectively.
Figure 7 shows specificity of CNN during disease
detection and grading of tobacco is greater than that of RNN REFERENCES
and DNN because the CNN is a general algorithm that is very
flexible to learn almost everything including image [1]. Abu-Naser, S. S., Kashkash, K. A., & Fayyad, M.
classification. (2010). Developing an expert system for plant disease
diagnosis. Journal of Artificial Intelligence;
Scialert,3(4), 269-276.
specificity between algorithms [2]. Jiyu Peng, Kunlin Song “Fast detection of tobacco
mosaic virus infected tobacco using laser-induced
92.8 breakdown spectroscopy,” March 2017
92.6 [3]. Abu-Nasser, B. S., & Abu-Naser, S. S. (2018). Cognitive
Specificity

92.4 System for Helping Farmers in Diagnosing Watermelon

92.2 Diseases. International Journal of Academic Information
92 Systems Research (IJAISR), 2(7), 1-7.
91.8 [4]. Abu Naser, S. S., & Shaath, M. Z. (2016). Expert system
91.6 urination problems diagnosis. World Wide Journal of
DNN RNN CNN Multidisciplinary Research and Development, 2(5), 9-
19.
Algorithm
[5]. J. Liu, J. Shen, Z. Shen, and R. Liu, “Grading tobacco
leaves based on image processing and generalized
Fig 7:- specificity between algorithms regression neural network,” in Proc. International
Conference on Intelligent Control, Automatic Detection
and High-End Equipment, 2012, pp. 89-93.

IJISRT22SEP836 www.ijisrt.com 1124

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[6]. Y. Sari and R. A. Pramunendar, “Classification quality
of tobacco leaves as cigarette raw material based on
artificial neural networks,” Int. J. Comput. Trens
Technol., vol. 50, no. 3, pp. 147-150, August 2017.
[7]. A. Santhosh, K. Saravanakumar, P. Saravanakumar, and
S. Vimalraj, R. Saranya “DETECTION OF LEAF
DISEASES AND CLASSIFYING THEM USING
MULTICLASS SVM, ” March, 2019.
[8]. S. K. Dasari, K. R. Chintada, and M. Patruni, “Flue-
cured tobacco leaves classification: A generalized
approach using deep convolutional neural networks,”
Cognitive Science and Artificial Intelligence Advances
and Application, pp. 13-21, 2018.
[9]. Yi, S. C., Hajaligol, M. R., & Jeong, S. H. (2005). The
prediction of the effects of tobacco type on smoke
composition from the pyrolysis modeling of tobacco
shreds. Journal of analytical and applied pyrolysis, 74(1-
2), 181-192
[10]. W. S. Jeon and S. Y. Rhee, “Plant leaf recognition using
a convolutional neural network,” International Journal of
Fuzzy Logic and Intelligent Systems, vol. 17, no. 1, pp.
26-34, March 2017
[11]. S. Pan, M. Kudo, and J. Toyama, “Edge detection of
tobacco leaf images based on fuzzy mathematical
morphology,” in Proc. International Conference on
Information Science and Engineering, 2009, pp. 1219-
1222.
[12]. V. Premalatha, M. Sumithra, S. Deepak, and P.
Rajeswari, “Implementation of spatial fem for leaf image
segmentation in pest detection,” Int. J. Adv. Res.
Comput. Sci. Softw. Eng., vol. 4, no. 10, pp. 471-477,
October 2014.
[13]. P. Viola and M. Jones, “Rapid object detection using a
boosted cascade of simple features,” in Proc. the IEEE
Computer Society Conf. on Computer Vision and
Pattern Recognition, 2001, p. 1.
[14]. Mohanad H. Al-Qadi, Mohammed F. El-Habibi
“Developing an Expert System to Diagnose Tomato
Diseases,” Pages: 34-40, May 2022
[15]. Abu Naser, S. S., & Zaqout, I. S. (2016).
Knowledgebased systems that determine the appropriate
students major: In the faculty of engineering and
information technology. World Wide Journal of
Multidisciplinary Research and Development, 2(10), 26-
34.

IJISRT22SEP836 www.ijisrt.com 1125