SRI KRISHNA COLLEGE OF TECHNOLOGY

An Autonomous Institution | Accredited by NAAC with 'A' Grade

Affiliated to Anna University | Approved by AICTE
KOVAIPUDUR, COIMBATORE 641042

PLANT DISEASE CLASSIFICATION AND

DETECTION

AN AIML PROJECT REPORT

Submitted by

THARUNIYAA J M Register No: 727821TUCS244

SANJAY S Register No: 727821TUCS216

NAVEENRAJ M Register No: 727822TUCS801

in partial fulfillment for the award of the

degree of

BACHELOR OF

ENGINEERING IN

COMPUTER SCIENCE AND ENGINEERING

OCTOBER 2024
 SRI KRISHNA COLLEGE OF TECHNOLOGY
An Autonomous Institution | Accredited by NAAC with 'A' Grade
Affiliated to Anna University | Approved by AICTE
KOVAIPUDUR, COIMBATORE 641042

BONAFIDE CERTIFICATE

Certified that this project report “PLANT DISEASE CLASSIFICATION AND

DETECTION” is the bonafide work of “THARUNIYAA J M, SANJAY S,
NAVEENRAJ M” who carried out the AIML project work under my supervision.

SIGNATURE SIGNATURE
Dr. K.A. JAYABALAJI Dr. M.UDHAYAMOORTHI
SUPERVISOR HEAD OF THE DEPARTMENT
Assistant Professor, Associate Professor,
Department of Computer Science Department of Computer Science
and Engineering and Engineering
Sri Krishna College of Technology, Sri Krishna College of
Technology, Coimbatore-641042. Coimbatore-641042.

Certified that the candidates were examined by us in the AIML Project Viva Voce
examination held on at Sri Krishna College of Technology,
Kovaipudur, Coimbatore -641042

INTERNAL EXAMINER EXTERNAL EXAMINER

 III

ABSTRACT

Plant diseases pose a significant challenge to global food security, with the potential to
cause severe crop losses and economic hardships, making early detection critical for
effective disease management. In response to this challenge, this research presents a
novel Convolutional Neural Network (CNN) algorithm specifically designed to
identify plant diseases through image-based symptom recognition. Recent
advancements in image recognition have been leveraged to develop a system that is
both accurate and efficient. To train the model, a comprehensive dataset was compiled,
featuring a diverse range of plant species and disease symptoms. These images
underwent pre-processing steps to standardize quality, which helped enhance the
model’s generalization abilities across different contexts. The CNN was trained using
a stochastic gradient descent algorithm, with extensive hyperparameter tuning to
prevent overfitting and ensure model robustness. The model's performance was
rigorously tested on an independent dataset, with metrics such as accuracy, precision,
recall, F1 score, and a confusion matrix used to evaluate its effectiveness. Beyond
simply identifying plant diseases, the system offers actionable insights to users, such
as tailored fertilizer recommendations based on the diagnosed disease. This additional
functionality promotes healthier crop management practices and aligns with
sustainable agriculture efforts. The system’s ability to provide early and precise
disease detection, coupled with its recommendations, represents a valuable tool for
improving agricultural productivity and contributing to global food security by helping
mitigate the impacts of plant diseases. Through further development and refinement,
this approach has the potential to significantly aid farmers and agricultural
professionals in better managing plant health.
 IV

ACKNOWLEDGEMENT

First and foremost we thank the Almighty for being our light and for
showering his gracious blessings throughout the course of this project.
We express our gratitude to our beloved Principal, Dr. M.G.
Sumithra, for providing all facilities.
We are grateful to our beloved Dean, Computing Sciences Dr. T.
Senthilnathan, for his tireless and relentless support.
With a grateful heart, our sincere thanks to our Head of the
Department Dr. M. Udhayamoorthi, Department of Computer Science and
Engineering for the motivation and all support to complete the project work.
We thank Ms.R . G N A N A K U M A R I , Department of Computer
Science and
Engineering, Project Coordinator for her motivation and support
We extend our sincere thanks to our project supervisor
Dr.K.A.JAYABALAJI , Assisstant Professor , Department of Computer
Science and Engineering, for the valuable guidance and suggestions in all
aspects that aided us in ameliorating our skills.
We are thankful to all the Teaching and Non-Teaching Staff of
the Department of Computer Science and Engineering and to all those who
have directly and indirectly extended their help to us in completing this
project work successfully.
We extend our sincere thanks to our family members and our beloved
friends, who had been strongly supporting us in all our endeavors.
 V

TABLE CONTENTS
CHAPTER TITLE PAGE NO.
NO.

ABSTRACT III

ACKNOWLEDGEMENT IV

LIST OF FIGURES VII

LIST OF ABREVIATION VIII

1. INTRODUCTION 1

2. LITERATURE REVIEW 2

3. DESIGN METHODOLOGY 3

3.1 EXISTING SYSTEM 3

3.1.1 DRAWBACKS 4

3.2 PROPOSED SYSTEM 6

3.2.1 ADVANTAGES 6

3.2.2 ARCHITECTURE DIAGRAM 7

3.3 MODULE DESCRIPTION 8

4. IMPLEMENTATION AND RESULTS 11

5. CONCLUSION AND FUTURE WORK 15

5.1 CONCLUSION 15

5.2 FUTURE WORK 16

REFERENCES 17
 VI

LIST OF FIGURES

FIGURE NO. TITLE PAGE NO.

3.1 ARCHITECTURE DIAGRAM 7

4.1 DATA LOADING 12

4.2 IMAGE UPLOAD 13

4.3 MODEL PREDICTION 14

 VII

LIST OF ABBREVIATIONS

ACRONYM ABBREVIATION
KNN K-Nearest Neighbor Algorithm
CNN Convolutional Neural Network
RNN Recurrent Neural Network
DNN Deep Neural Network
 1

CHAPTER 1

INTRODUCTIO

Plant disease detection and classification systems are rapidly becoming essential tools in
agriculture, influencing many aspects of crop management. From precision farming to
large-scale monitoring, this technology is being adopted across different agricultural
sectors. Its ability to recognize and classify plant diseases based on visual symptoms has
made it indispensable in maintaining crop health and yield. However, its increasing use
in critical decision-making processes, such as disease management, pest control, and
harvest predictions, raises significant concerns. When a plant disease detection system
is used to identify infections or recommend treatment actions, any mistakes or
inaccuracies can have serious repercussions for food security and agricultural
productivity. Thus, there is an urgent need to make these systems explainable,
especially when decisions impact crucial agricultural outcomes. Making plant disease
detection processes explainable involves offering clear, understandable insights into
how and why a system reaches a certain diagnosis. Traditional black-box models make
this difficult, as they lack transparency in their decision-making processes. A key aspect
of building trust in these systems is ensuring that farmers and agricultural decision-
makers can understand the reasoning behind the system's classification results.
Particularly in farming, stakeholders want to know why a specific disease was detected
or missed by the system, and how reliable that diagnosis is under various environmental
conditions. One of the greatest challenges for plant disease detection systems arises
when they are used in field environments. Unlike controlled environments like
greenhouses, where factors such as lighting, soil quality, and plant positioning are
managed, field environments introduce variability that complicates the classification
task. Changes in weather (rain, sunlight, fog), soil conditions, and occlusions
(overlapping leaves, dirt) all interfere with the system’s ability to accurately capture and
analyze plant features.
 2

CHAPTER 2

LITERATURE

REVIEW

PLANT DISEASE DETECTION

Plant Disease Detection and Classification is a rapidly emerging field aimed at making
the inner workings of plant disease detection models understandable to humans. The
need for explainability has grown as these systems have become more integrated into
critical applications, such as precision agriculture and crop health monitoring. Initial
efforts in explainable plant disease detection focused on improving the interpretability
of plant disease representations. Singh et al. (2019) introduced an image segmentation
approach that highlighted specific diseased regions in plants, enhancing the
interpretability of the disease detection process. The resulting representations were more
structured, with each segment corresponding to a particular diseased part of the plant.
Patel et al. (2020) employed Class Activation Maps (CAMs) to identify the most
significant leaf regions contributing to the classification tasks, providing insights into
which parts of the plant image were most relevant for detecting specific

DEEP LEARNING FOR IMAGE-BASED PLANT DISEASE DETECTION

Sharada Prasanna Mohanty et al. have proposed a system addressing the significant
threat that crop diseases pose to food security, particularly in regions lacking adequate
infrastructure for rapid identification. The combination of increasing global smartphone
penetration and recent advancements in computer vision, driven by deep learning, has
enabled the development of smartphone-assisted disease diagnosis. In this system, a
public dataset containing 54,306 images of diseased and healthy plant leaves, collected
under controlled conditions, was used to train a deep convolutional neural network.
This model is capable of identifying 14 crop species and diagnosing 26 diseases
(including the absence of disease), offering a practical solution for real-time plant
disease detection and helping farmers monitor crop health effectively.
 3

CHAPTER 3

DESIGN

METHODOLOGY

3.1 EXISTING SYSTEM

A Plant disease diagnosis is crucial for agriculture due to its role in increasing crop
production. Recent advancements in image processing provide new solutions through
visual plant disease analysis. However, there has been limited research in this area, and
systematic investigations are even rarer. In this paper, we systematically explore the
problem of visual plant disease recognition for diagnosis. Unlike typical images, plant
disease images present unique challenges due to randomly distributed lesions, diverse
symptoms, and complex backgrounds, making it difficult to capture discriminative
information. To advance plant disease recognition research, we construct a new large-
scale dataset containing 271 plant disease categories and 220,592 images. Using this
dataset, we address plant disease recognition by reweighting both visual regions and
losses to emphasize the diseased parts of plants. First, we calculate the weights of all
divided image patches based on their cluster distribution, indicating the discriminative
level of each patch. Then, we assign a weight to each loss for each patch-label pair
during weakly-supervised training, enabling the network to learn the most relevant
diseased regions. Finally, we extract patch features from the network trained with
reweighted losses and utilize an LSTM network to encode the weighed patch feature
sequences into a comprehensive feature representation.
 4

3.1.1 DRAWBACKS OF EXISTING SYSTEM

The advent of machine learning and deep learning has revolutionized the field of plant
disease detection, promising significant advancements in agricultural productivity and
sustainability. However, despite the remarkable potential of these methods, there are
substantial challenges that hinder their widespread implementation. One of the most
critical obstacles is the need for a large-scale plant disease dataset to train the network
effectively. A robust dataset is essential to ensure that the model learns to differentiate
between various plant diseases accurately. Collecting a comprehensive dataset can be
particularly challenging for rare or newly emerging diseases, which may not have
sufficient samples available for analysis. This scarcity of data not only limits the
training process but also risks overfitting the model to specific cases, leading to poor
generalization when faced with novel examples in real-world scenarios. Moreover,
obtaining high-quality label data is time-consuming and resource-intensive, requiring
the expertise of trained agronomists or plant pathologists to ensure accuracy, further
complicating the dataset acquisition process.

Another significant challenge associated with the current methods is their computational
expense. Advanced algorithms often require a considerable amount of computational
power, especially when handling high-resolution images of plants. This process
typically involves dividing each image into smaller patches and computing the weights
of each patch to assess its contribution to the overall health of the plant. Such operations
can be intensive, demanding significant processing time and resources, which may not
be feasible for smaller agricultural operations or regions with limited access to
technological infrastructure. Consequently, the high computational demands can restrict
the practical application of these methods, particularly in areas where farmers need
quick, real-time diagnostics to manage crop health effectively. Furthermore, the reliance
on sophisticated hardware can also exacerbate issues related to scalability and
accessibility, creating barriers for widespread adoption in low-resource settings. In
addition to the data and computational challenges, the existing methods often struggle
with robustness against variations in environmental conditions, such as lighting,
background, and weather factors.
 5

Natural environments are inherently variable, and fluctuations in light intensity can
significantly impact image quality and the visibility of disease symptoms on plants. For
instance, shadows and reflections can obscure important features, leading to
misclassification or missed detections. Similarly, diverse backgrounds—ranging from
soil to different types of vegetation—can introduce noise into the images, complicating
the task of identifying the disease. To enhance the robustness of these models, it is
imperative to develop techniques that can effectively normalize these variations,
ensuring accurate detection regardless of the environmental context. Failure to address
this issue can limit the applicability of these methods, rendering them ineffective in real-
world agricultural settings where conditions are far from controlled.

Furthermore, the current state of plant disease detection methods often falls short in
their ability to identify all types of plant diseases. Many algorithms are trained on
specific diseases prevalent in certain regions or crops, leaving gaps in their detection
capabilities for other diseases that may not be represented in the training data. This
limitation poses a significant risk to agricultural practices, as farmers may remain
unaware of emerging threats or may be unable to respond effectively to a disease
outbreak. Consequently, a more inclusive approach to dataset curation is needed, one
that encompasses a wider variety of diseases, plant species, and environmental
conditions. This would not only enhance the model's training but also provide a more
comprehensive understanding of plant health management.

In summary, while machine learning and deep learning hold great promise for plant
disease detection, several formidable challenges remain. The necessity for large-scale,
high-quality datasets poses significant barriers, particularly for rare diseases.
Additionally, the computational demands of current methodologies can limit their
feasibility for practical use, especially in resource-limited settings. Moreover, existing
methods often lack robustness against environmental variability and are not
comprehensive enough to identify all potential plant diseases. Addressing these
challenges will be crucial for advancing the field and realizing the full potential of
technology in enhancing agricultural productivity and sustainability
 6

3.2 PROPOSED SYSTEM

The plant disease detection system boasts a user-friendly interface designed to simplify
the process for users, enabling them to effortlessly upload images of their plants for
analysis. This intuitive interface is crucial, as it accommodates users of varying
technical backgrounds, ensuring accessibility for both novice and experienced gardeners
alike. Once images are uploaded, the system undertakes a pre-processing phase to
standardize the size and normalize pixel values, which is essential for maintaining
compatibility with the Convolutional Neural Network (CNN) model employed in the
analysis. Standardization and normalization are critical steps that enhance the model’s
accuracy and reliability, as they ensure that the input data remains consistent, allowing
the CNN to focus on detecting the unique visual symptoms indicative of specific plant
diseases. After the pre-processing is completed, the trained CNN meticulously analyzes
the images to identify potential plant diseases. This advanced neural network is
designed to recognize subtle patterns and features that may be indicative of various
health issues, effectively translating visual symptoms into actionable insights. The
power of CNNs lies in their ability to learn from vast amounts of data, allowing them to
improve over time and become increasingly adept at distinguishing between different
diseases. Upon successfully identifying a disease, the system goes a step further by
providing tailored fertilizer recommendations. These suggestions are tailored to address
the specific issues identified, enabling users to take immediate action to mitigate the
impact of the disease and promote overall plant health. The system presents the analysis
results, including disease predictions and fertilizer recommendations, in a clear and
comprehensible manner. This transparency is vital, as it empowers users to make
informed decisions regarding their plant care practices. By visually displaying the
results, the system enhances user understanding, allowing individuals to recognize the
symptoms and comprehend the rationale behind the recommended treatments. Such an
approach not only fosters greater engagement with the technology but also encourages
users to take proactive measures in maintaining their plants' health. Ultimately, this
comprehensive system, with its focus on user experience and effective analysis, aims to
revolutionize plant care, ensuring that individuals can easily identify and address plant
diseases while promoting healthier growth.
 7

3.3 ARCHITECTURE STRUCTURE

Plant
Load Pre- diseases Result
dataset processing using CNN analysis
 8

Figure 3.1 ARCHITECTURE DIAGRAM

 9

3.4 MODULE DESCRIPTION

The Plant Disease Detection and Classification System is an innovative solution

designed to identify and classify plant diseases through image analysis. Using advanced
image processing techniques and a Convolutional Neural Network (CNN), the system
analyzes high-resolution images of plant leaves to detect visual symptoms indicative of
various diseases. After pre-processing the images for optimal quality, the CNN
accurately recognizes patterns associated with specific health issues. Upon identifying a
disease, the system provides actionable recommendations for treatment, such as tailored
fertilizer or pesticide suggestions, empowering users to take informed actions to
maintain healthy plants and enhance agricultural productivity.

1. Image Upload Module

Purpose: To enable users to easily upload images of their plants for analysis through a
user-friendly interface.

Functionality: This module acts as the primary interface for user interaction, allowing
for effortless image uploads. It efficiently manages file selection and performs validation
checks to ensure that the images conform to the required format and size specifications.
By ensuring that only valid images are processed, the module minimizes errors and
enhances the reliability of the system. Once the upload is successful, the images are
temporarily stored, paving the way for further analysis by the plant disease detection
algorithms.

2. Image Pre-processing Module

Purpose: To prepare uploaded images for analysis by transforming them into a format
compatible with the Convolutional Neural Network (CNN).

Functionality: Upon receiving the uploaded image, the Image Pre-processing Module
performs essential transformations to ensure the image is suitable for analysis.
 10

It resizes the image to the required dimensions, such as 128x128 pixels, and normalizes
the pixel values to a range of 0 to 1 to standardize the input data. Additionally, it converts
the image into an appropriate format for seamless integration with the CNN model. This
pre-processing step is critical, as it enhances the model's ability to accurately interpret the
image data, ultimately improving the reliability of disease detection and classification.

3. Disease Prediction Module

Purpose: To serve as the system's core component, leveraging a trained Convolutional

Neural Network (CNN) to analyze pre-processed images for plant disease identification.

Functionality: This module is responsible for utilizing the trained CNN model to assess
the pre-processed image and predict the presence of specific plant diseases based on
visual symptoms. It generates an output that includes the predicted disease class and a
confidence score, offering valuable insights into the plant's health status. By providing a
clear indication of disease likelihood, the module enables users to make informed
decisions regarding timely interventions and treatment strategies. Ultimately, this
functionality enhances the system's effectiveness in supporting plant health management
and improving agricultural outcomes.

4. Fertilizer Recommendation Module

Purpose: Provide tailored fertilizer suggestions to help mitigate the effects of identified
plant diseases.

Functionality: Upon disease identification, this module generates customized fertilizer

recommendations based on predefined logic or a comprehensive database linking
various disease types with suitable fertilizer options and application methods. It takes
into account the specific needs of the affected plants, ensuring that the suggested
fertilizers address the nutritional deficiencies caused by the disease. By delivering these
targeted recommendations, the module empowers users to effectively manage their
plants' health and promote recovery,
 11

ultimately enhancing overall agricultural productivity. This functionality ensures that

users can take informed actions to support their plants in overcoming disease-related
challenges. Ultimately, this functionality enhances the system's effectiveness in
supporting plant health management and improving agricultural outcomes.

5. Result Display Module

Purpose: To present analysis results, including predicted diseases and corresponding

fertilizer recommendations, in a clear and informative format for the user.

Functionality: This module plays a crucial role in communicating the outcomes of the
analysis, displaying the predicted disease name alongside tailored fertilizer
recommendations to facilitate effective plant management. In addition to textual
information, the module may incorporate visual aids, such as images or links to
supplementary resources, providing users with a more comprehensive understanding of
the identified disease and its treatment options. By ensuring that users have access to
detailed insights and care tips, the Result Display Module enhances the overall user
experience, empowering them to make informed decisions about their plants' health.
Ultimately, this functionality fosters better engagement and encourages proactive
measures in plant care and disease management.
 12

CHAPTER 4
IMPLEMENTATION AND
RESULTS

The implementation of the Plant Disease Detection System involves several key steps to
ensure accurate diagnosis and effective user interaction. The process begins with
hardware setup, which includes selecting appropriate devices such as smartphones or
cameras for capturing high-resolution images of plant leaves. Once the hardware is in
place, the next step is to develop the software components. This includes setting up the
programming environment, typically using Python, and integrating libraries such as
OpenCV for image processing and TensorFlow for building and training the
Convolutional Neural Network (CNN) model. The next phase involves data collection
and preparation, where a comprehensive dataset of plant images representing various
diseases is gathered. This dataset is pre-processed to standardize image dimensions and
normalize pixel values, ensuring compatibility with the CNN model. After preparing the
data, the system proceeds to train the CNN model using the prepared dataset. This
involves feeding the model a significant number of labeled images, allowing it to learn
and recognize patterns associated with different plant diseases. Techniques such as data
augmentation may be employed to enhance the robustness of the model by artificially
increasing the dataset's diversity. Once the model is trained, it is integrated into the
system for real-time analysis. The system includes an Image Upload Module for users to
upload images, followed by an image Pre-processing Module to prepare the images for
analysis. The Disease Prediction Module utilizes the trained CNN to analyze the pre-
processed images and predict the presence of diseases, outputting the predicted disease
class and confidence score. Finally, the Fertilizer Recommendation Module offers
tailored suggestions for managing identified diseases. To enhance user experience, a
**Result Display Module** showcases the analysis results in a clear format, providing
users with disease predictions and actionable recommendations.
 13

Figure 4.1 DATA LOADING

 14

Figure 4.2 IMAGE UPLOAD

 15

RESULT

Figure 4.3 DISEASE PREDICTION

 16

CHAPTER-5
CONCLUSION AND FUTURE WORK

5.1 CONCLUSION

The proposed Plant Disease Detection System significantly enhances the

accuracy of plant disease identification, offering a valuable tool for farmers aiming
to implement more effective management practices. By leveraging advanced image
processing and machine learning algorithms, the system provides actionable
insights that enable users to make informed decisions about their crops, ultimately
promoting healthier plant growth. With the ability to identify specific diseases and
suggest targeted fertilizer recommendations, the application empowers farmers to
respond swiftly and effectively to plant health issues. This capability not only
supports the immediate needs of users but also aligns with sustainable agricultural
practices by helping to mitigate the environmental impact of excessive fertilizer use.
By promoting optimal nutrient application, the system encourages healthier crops
and minimizes waste, contributing to more responsible farming methods.
Additionally, ongoing improvements and expansions of the model will enhance its
utility, making it applicable to a wider range of crops and diseases. By continually
updating the training dataset and refining the algorithms, the system can evolve
alongside the challenges faced by modern agriculture. This innovative approach not
only addresses the pressing issue of plant disease management but also contributes
to overall agricultural productivity and food security. By providing farmers with the
tools they need to effectively combat plant health issues, the system plays a crucial
role in tackling critical challenges in global agriculture, ensuring that food
production can keep pace with growing demands while maintaining environmental
integrity. Ultimately, the integration of this technology in agricultural practices
represents a significant step forward in the quest for sustainable and productive
farming solutions, benefiting both farmers and the broader ecosystem.
 17

5.2 FUTURE WORK

The future development of the Plant Disease Detection System aims to address
critical areas that will significantly enhance its effectiveness and usability for
farmers and agricultural professionals. One of the primary objectives is to
incorporate additional plant species and diseases into the training dataset. By
expanding the dataset, the model's generalization and accuracy can be improved
across diverse agricultural contexts, ensuring that users can rely on the system for a
broader range of plant health issues. Furthermore, creating a mobile version of the
system is essential for enhancing accessibility, allowing users to conveniently
access disease identification and recommendations while in the field. A mobile
application would empower farmers to quickly capture images of their plants,
receive instant feedback, and make timely decisions to manage their crops
effectively. Additionally, the integration of IoT devices for real-time monitoring of
plant health represents a transformative opportunity. By utilizing sensors that
measure environmental factors such as soil moisture, temperature, and humidity, the
system can provide dynamic adjustments to care and management practices,
allowing farmers to respond proactively to changing conditions. Refining the
fertilizer recommendation module is another crucial aspect of future work. By
considering localized agricultural practices, soil health, and climatic factors, the
system can offer more relevant and tailored fertilizer suggestions, enhancing the
effectiveness of nutrient application and promoting sustainable farming practices.
Finally, incorporating user feedback mechanisms will be vital for the ongoing
improvement of the system. By allowing users to provide insights into the accuracy
of disease predictions and the effectiveness of recommended treatments, developers
can continually refine the system to better meet the evolving needs of farmers.
Collectively, these initiatives will not only enhance the functionality and reliability
of the Plant Disease Detection System but also contribute to improved agricultural
productivity and sustainability, addressing the pressing challenges faced by the
agricultural sector today.
 18

REFERENCES

1. V. Singh and A. K. Misra, "Detection of plant leaf diseases using image

segmentation and soft computing techniques," Information Processing in
Agriculture, vol. 4, pp. 41–49, Mar. 2017.

2. S. H. Lee, C. S. Chan, S. J. Mayo, and P. Remagnino, "How deep learning

extracts and learns leaf features for plant classification," Pattern Recognition,
vol. 71, pp. 1–13, Nov. 2017.

3. G. Farjon, O. Krikeb, A. B. Hillel, and V. Alchanatis, "Detection and counting

of flowers on apple trees for better chemical thinning decisions," Precision
Agriculture, vol. 21, pp. 1–19, Aug. 2019.

4. S. P. Mohanty, D. P. Hughes, and M. Salathé, "Using deep learning for image-

based plant disease detection," Frontiers in Plant Science, vol. 7, p. 1419, Sep.
2016.

5. F. Ren, W. Liu, and G. Wu, "Feature reuse residual networks for insect pest
recognition," IEEE Access, vol. 7, pp. 122758–122768, 2019.