A Snake Classification Model for Snakebite Envenoming Management 

 

 

 

Mariam Mabinda 

136553 

 

 

 

 

 

Masters of Science in Information Technology  

2022



 

A Snake Classification Model for Snakebite Envenoming Management 

 

 

 

 Mariam Mabinda 

136553 

 

 

Submitted in partial fulfillment of the requirements for the Degree of Masters of 

Science in Information Technology at Strathmore University. 

 

 

School of Computing and Engineering Sciences 

Strathmore University 

Nairobi, Kenya 

 

 

October, 2022 

This thesis is available for Library use on the understanding that it is copyright 

material and that no quotation from the thesis may be published without proper 

acknowledgement. 



 
 

ii 

 

Declaration and Approval 

Declaration 

I, Mariam Mabinda declare that this work has not been previously submitted and 

approved for the award of a degree by this or any other University. To the best of my 

knowledge and belief, the thesis contains no material previously published or written 

by another person except where due reference is made in the thesis itself. 

© No part of this thesis may be reproduced without the permission of the author and 

Strathmore University  

Mariam Mabinda 

Sign: ________________________  

Date: 13/7/2022 

Approval 

The thesis of Mariamu Mabinda was reviewed and approved by the following: 

 

Dr. Dickson Owuor, 

Senior Lecturer, School of Computing & Engineering Sciences, 

Strathmore University. 

 

Dr. Julius Butime, 

Dean, School of Computing & Engineering Sciences, 

Strathmore University. 

 

Dr. Bernard Shibwabo, 

Director of Graduate Studies, 

Strathmore University. 

  



 
 

iii 

 

Abstract 

Snakebite envenoming is a potentially life-threatening disease caused by the injection 

of toxins through a bite or venom sprayed into the victim’s eyes by certain venomous 

snake species. WHO program dubbed Neglected Tropical Disease Program (NTD) of 

2019 indicated that about 5.4 million snake bites occur each year, resulting in 1.8 to 

2.7 million cases of envenoming. Of this about 81,000–138,000 deaths occur and 

approximately 400,000 people are permanently disabled annually. Kenya is 

approximated to have more than 15,000 bites annually. Correct identification of the 

snake species in question plays a critical role in the proper administration of the right 

first aid and suitable prescription of the anti-venom for the patient. Currently, there is 

no automated method of identifying snake species using images in Kenya. The usual 

practice is to, kill the snake and carry it along with the patient to the hospital or to give 

a visual feature description of the biting snake. Also, a blood test can be done to look 

for the presence of toxins associated with the described snake species. The challenge 

however is, that the time required for test results to be out can jeopardize patients' 

survival depending on the type of venom injected. Furthermore, the cost associated 

with the test is also punitive. The ability to correctly classify a snake species is a 

challenging task for both humans and machines mainly because of subtle differences 

between different snake species and strong variety within the same species. Existing 

studies used a combination of feature extraction methods and deep neural networks 

and yielded an accuracy of 90%. These models applied Principal component analysis 

(PCA) and Linear discriminant analysis (LDA) as feature extractors. However, the use 

of the Singular Value Decomposition (SVD) algorithm was not explored despite its 

apparent advantage. This research study solved the classification challenge by creating 

a Kenyan snake species dataset and developing a snake species classification model 

that predicts a snake species based on the image and classifies it according to its venom 

toxicity. The study carried out feature reduction of the images using the SVD algorithm 

and passed these features as input to a deep learning model using transfer learning. 

The model was trained using 4521 labelled snake images via supervised transfer 

learning using MobileNetV2. The model was trained, validated, tested, and achieved 

an outstanding classification accuracy of 96 %. The model surpassed the accuracy of 

the existing model. 



 
 

iv 

 

Keywords: Singular Value Decomposition, snake classification, feature reduction, 

transfer learning with convolutional neural networks. 

  



 
 

v 

 

Definition of Terms 

Classification Model  A term that describes drawing a conclusion 

from an input value given for training and can 

predict the category for new data (Sarker et al., 

2019) 

Neural Networks This consists of thousands or even millions of 

simple processing nodes that are densely 

interconnected. They are a means of machine 

learning in which a computer learns to perform 

a task by analyzing specific examples (Ian et al., 

2016) 

Snakebite Envenomation Injection of snake venom by venomous snake 

species to a human through a bite or spraying the 

venom into the eyes mostly for defense purposes  

(WHO, 2019). 

Transfer Learning Transfer learning is the use of stored knowledge 

gained from a pre-trained neural network model, 

to train a different but related neural network 

model with a smaller dataset (Perera & Patel, 

2019). 

Data Augmentation The process of artificially inflating datasets 

using a series of transformations such as 

rotation, cropping, flipping (Taylor and 

Nitschke, 2017). 

 Deep Learning This is a subfield of machine learning concerned 

with the development of neural networks that are 

deeper or consist of several convolutions. The 

deeper learning enables them to learn and model 

complex scenarios (Nielsen, 2015). 

 



 
 

vi 

 

Table of Contents 

Declaration and Approval ............................................................................................ ii 

Abstract ....................................................................................................................... iii 

Definition of Terms ...................................................................................................... v 

Table of Contents ........................................................................................................ vi 

List of Abbreviations................................................................................................... xi 

List of Tables.............................................................................................................. xii 

List of Figures ........................................................................................................... xiii 

Acknowledgments ..................................................................................................... xiv 

Dedication .................................................................................................................. xv 

Chapter 1: Introduction ................................................................................................ 1 

1.1 Background ........................................................................................................ 1 

1.2 Problem Statement ............................................................................................. 3 

1.3 Objectives ........................................................................................................... 4 

1.3.1 General Objective ........................................................................................ 4 

1.3.2 Specific Objectives ....................................................................................... 5 

1.4 Research Questions ............................................................................................ 5 

1.5 Justification ........................................................................................................ 5 

1.6 Scope and Limitation .......................................................................................... 6 

Chapter 2: Literature Review ....................................................................................... 6 

2.1 Introduction ........................................................................................................ 6 

2.2 Snakebite Envenomation Overview ................................................................... 6 

2.2.1 Traditional methods of Identifying snake .................................................... 7 

2.2.2 Challenges of Snakebite Envenomation Management ................................. 9 

2.3 Application of Technology in Snakebite Envenomation Management ............ 10 

2.3.1Use Natural Language Processing (NLP) .................................................. 10 

2.3.2 Use of Image Classifier ............................................................................. 11 

2.4 Steps of Classifying Image ............................................................................... 13 

2.4.1 Image Pre-Processing................................................................................ 13 



 
 

vii 

 

2.4.2 Feature Extraction. .................................................................................... 14 

2.4.3 Image Classification .................................................................................. 14 

2.5 Feature Extraction Techniques ......................................................................... 14 

2.5.1 Principal Component Analysis (PCA) ....................................................... 14 

2.5.2 Linear Discriminant Analysis (LDA) ......................................................... 15 

2.5.3 Singular Value Decomposition (SVD) ....................................................... 15 

2.6 Machine Learning Techniques used in Image Classification ........................... 16 

2.6.1 Support Vector Machine ............................................................................ 16 

2.6.2 K-Nearest Neighbor (KNN) ....................................................................... 17 

2.6.3 Convolution Neural Network (CNN) ......................................................... 17 

2.6.4 Transfer Learning ...................................................................................... 20 

2.6.5 Tensor Flow Framework ........................................................................... 21 

2.6.6 Tensor Flow Lite. ....................................................................................... 21 

2.7 Research Gap .................................................................................................... 22 

2.8 Conceptual Framework .................................................................................... 23 

Chapter 3: Research Methodology ............................................................................. 25 

3.1 Introduction ...................................................................................................... 25 

3.2 Research Design ............................................................................................... 25 

3.3 Population and Sampling .................................................................................. 25 

3.3.1. Population ................................................................................................. 25 

3.3.2 Sampling .................................................................................................... 26 

3.4 Data Collection ................................................................................................. 26 

3.5 Image Pre-Processing and Analysis ................................................................. 27 

3.5.1 Image Augmentation .................................................................................. 27 

3.5.2 Standardization .......................................................................................... 27 

3.5.3 Folder Structure ......................................................................................... 27 

3.6 System Development Methodology ................................................................. 27 

3.6.1 Agile System Development ......................................................................... 27 

3.6.2 System Analysis .......................................................................................... 28 

3.6.3 System Design ............................................................................................ 28 

3.6.4 System Implementation .............................................................................. 28 



 
 

viii 

 

3.7 Research Quality .............................................................................................. 29 

3.7.1 Research Validity ....................................................................................... 30 

3.7.2 Reliability ................................................................................................... 30 

3.8 Ethical Considerations ...................................................................................... 30 

Chapter 4: System Analysis, Design, and Architecture ............................................. 31 

4.1 Introduction ...................................................................................................... 31 

4.2 Requirements Gathering and Analysis ............................................................. 31 

4.2.1 Functional Requirements ........................................................................... 31 

4.2.2Non-Functional Requirements .................................................................... 32 

4.3 System Design .................................................................................................. 32 

4.3.1 System Architecture ................................................................................... 32 

4.3.2 Use-Case Diagram and Descriptions ........................................................ 33 

4.3.3 System Sequence Diagram ......................................................................... 37 

4.3.4 Entity Relationship Diagram ..................................................................... 38 

4.3.5 Class Diagram ........................................................................................... 39 

4.3.6 Wireframes of the Mobile Application ....................................................... 40 

Chapter 5: Model Development, Implementation, and Testing ................................. 43 

5.1 Introduction ...................................................................................................... 43 

5.2 Development environment and Language ........................................................ 43 

5.2.1 Software Requirements .............................................................................. 43 

5.2.2. Hardware Requirements ........................................................................... 44 

5.3 Model Components .......................................................................................... 45 

5.3.1 Storage ....................................................................................................... 46 

5.3.2 Input Layer ................................................................................................. 46 

5.3.3 Hidden Layer ............................................................................................. 47 

5.3.4 Output Layer .............................................................................................. 48 

5.4 Model Development ......................................................................................... 48 

5.4.1 Image Dataset Collection .......................................................................... 48 

5.4.2 Import of necessary Library ...................................................................... 49 

5.4.3 Fetch the image .......................................................................................... 49 

5.4.4 Data split.................................................................................................... 50 



 
 

ix 

 

5.4.5 Feature Reduction...................................................................................... 50 

5.4.6 Image Pre-processing and Augmentation.................................................. 51 

5.5 MobileNetV2 Retraining .................................................................................. 52 

5.6 Hyperparameter tuning ..................................................................................... 54 

5.7 Model Validation .............................................................................................. 54 

5.8 Converting Tensor flow Model to Tensor flow Lite ........................................ 56 

5.9 Android Mobile Application Development ...................................................... 56 

5.9.1 Front End Development ............................................................................. 56 

5.9.2 Login and Registration .............................................................................. 56 

5.9.3 Choosing an Image .................................................................................... 57 

5.9.4 Image Classification .................................................................................. 58 

5.9.5 Viewing Antidote and First Aid Guidelines ............................................... 59 

5.10 Model Testing ................................................................................................. 61 

Chapter 6: Discussions ............................................................................................... 63 

6.1 Introduction ...................................................................................................... 63 

6.2 Model Validation .............................................................................................. 63 

6.2.1Precision ..................................................................................................... 64 

6.2.2 Recall ......................................................................................................... 65 

6.2.3 F1-Score..................................................................................................... 65 

6.3 Classification Report ........................................................................................ 67 

6.4 Research Contribution ...................................................................................... 68 

6.5 First Aid Details ............................................................................................... 69 

6.6 Accomplishment of the Objectives .................................................................. 69 

6.7 Advantages of the Developed System to Current Systems .............................. 70 

6.8 Shortfalls of the Research ................................................................................. 71 

6.9 Challenges Encountered ................................................................................... 71 

Chapter 7: Conclusion and Recommendation ............................................................ 73 

7.1 Conclusions ...................................................................................................... 73 

7.2 Recommendations ............................................................................................ 73 

7.3 Future Research Direction ................................................................................ 74 

References .................................................................................................................. 75 



 
 

x 

 

Appendices ................................................................................................................. 79 

Appendix A: MobileNet Model ............................................................................. 79 

Appendix B: Model Summary................................................................................ 80 

Appendix C: Training process ................................................................................ 86 

Appendix D: Confusion matrix and Classification Report..................................... 88 

Appendix E: Ethical Approval ............................................................................... 89 

Appendix F: Similarity Test Report ....................................................................... 90 

Appendix F: Link with demonstration of the App ................................................. 91 

Appendix G: Sample Snake Images Collected ....................................................... 92 

  

  



 
 

xi 

 

List of Abbreviations 

CEDD  Color and Edge Directivity Descriptor 

CNN   Convolution Neural Networks 

CPU  Central Processing Unit 

ERD  Entity Relationship Diagram 

GPU  Graphics Processing Unit 

ICT  Information Communication Technology. 

ID3  Iterative Dichotomiser 3 

KNN  K-Nearest Neighbor 

LDA  Linear Discriminant Analysis  

MOH  Ministry of Health 

NLP  Natural Language Processing 

NTD  Neglected Tropical Disease  

OOAD Object-Oriented Analysis and Design 

PCA  Principal Component Analysis 

ReLU  Rectified Linear Unit  

SVD  Singular Value Decomposition 

SVM  Support Vector Machine 

TFLite Tensor Flow Lite 

UML  Unified Modeling Language 

VGG16 Visual Geometry Group 16   

WHO  World Health Organisation 

 



 
 

xii 

 

List of Tables  

Table 2. 1 Summary of Literature Reviewed ............................................................. 22 

 

Table 4. 1 Register and Login Description ................................................................ 34 

Table 4. 2Load an Image and Classify Description ................................................... 35 

Table 4. 3 Query and View Results Description ........................................................ 35 

Table 4. 4 Recommend Antidote Description ............................................................ 36 

 

Table 5. 1 Software Requirements ............................................................................. 44 

Table 5. 2 Functional and Non-functional Testing. ................................................... 62 

  

 

 

 

 

  



 
 

xiii 

 

List of Figures  

Figure 2. 1: Image Classification Process (Nath et al., 2014) .................................... 13 

Figure 2. 2: CNN model Architecture (Progga et al., 2021) ...................................... 18 

Figure 2. 3 : Transfer Learning Architecture (Ribani & Marengoni, 2019) .............. 20 

Figure 2. 4 Conceptual Framework of the Snake Image classification Model .......... 24 

 

Figure 4. 1System Architecture.................................................................................. 33 

Figure 4. 2Use case Diagram ..................................................................................... 34 

Figure 4. 3 Login System Sequence Diagram ............................................................ 37 

Figure 4. 4 Classifying System Sequence Diagram ................................................... 38 

Figure 4. 5 Entity Relationship Diagram ................................................................... 39 

Figure 4. 6 Class Diagram .......................................................................................... 40 

Figure 4. 7 Login Wireframes .................................................................................... 41 

Figure 4. 8 Classification Wireframes ....................................................................... 42 

 

Figure 5. 1 Graphical Representation of the model ................................................... 45 

Figure 5. 2 Graphical Représentation of the input Layer ........................................... 46 

Figure 5. 3 Graphical Representation of the hidden layer ......................................... 47 

Figure 5. 4 Graphical representation of the output layer ........................................... 48 

Figure 5. 5 Collected Data in Excel File .................................................................... 49 

Figure 5. 6 Python code for mounting the colab to the drive. .................................... 50 

Figure 5. 7 Python code for performing SVD on train dataset .................................. 51 

Figure 5. 8 Python script for data augmentation ........................................................ 52 

Figure 5. 9 Model Epochs .......................................................................................... 53 

Figure 5. 10 Training accuracy vs validation accuracy.............................................. 55 

Figure 5. 11 Training Loss vs Validation Loss .......................................................... 55 

Figure 5. 12 Python script for converting the model to TFLite ................................. 56 

Figure 5. 13 Login and Registration Screenshot ........................................................ 57 

Figure 5. 14 Chose /select image Screenshot ............................................................. 58 

Figure 5. 15 Classification Screenshot. ...................................................................... 59 

Figure 5. 16 First Aid Guidelines ............................................................................... 60 

Figure 5. 17Antidote Information .............................................................................. 61 

 



 
 

xiv 

 

Acknowledgments 

First and foremost, I thank the Lord God Almighty for giving me the strength, time, 

resources, and good health as I pursued this course. It is by His grace and mercies that 

I have completed this research project. I thank my immediate supervisor, Dr. Dickson 

Ouwor for his relentless efforts, encouragement, and support to see that this research 

is completed successfully. Sincere gratitude to Dr. Vincent Omwenga for his guidance, 

encouragement, input, and support during the class seminars and throughout the 

research process. I also want to thank Dr. Bernard Shibwabo who guided me in the 

early stages of my research work. I would like to acknowledge and thank my 

classmates, MSIS class of 2022, and MSIT class of 2022 for the encouragement, 

knowledge sharing, and all the light moments we shared. The journey was worthwhile 

and the load was made lighter with your support. I acknowledge and appreciate you 

all. 

 

 

 

 

 

 

 

 

 

 

 

 

 

 



 
 

xv 

 

Dedication  

It is with genuine gratitude and warm regard that i dedicate this thesis to my family. A 

special dedication to my husband Eng. Justus Aufridus Otwani, who has been a 

constant source of support and encouragement always. I am blessed and thankful to 

have you as my friend and partner. To my sons Schneider Ikol and Schourn Ejakait, 

thank you for the support, encouragement, perseverance, and understanding as I was 

undertaking this research. You are my greatest cheerleaders and may the good Lord 

bless you all. A special dedication to my mum Franciscah Mwikali for being the lady 

I always want to emulate and for your prayers. In loving memory of my father, Albert 

Mabinda, who always pushed me to be the best version of me, as you would have said, 

“We did it!” On to the next one. To my sisters and brothers thanks for cheering me on 

and being there always when I needed you. 

 

 

 

 



 
 

1 

 

Chapter 1: Introduction 

1.1 Background  

Snakes are limbless carnivorous reptiles of the suborder Serpents (MOH-Kenya, 2019). 

They are cold-blooded, amniotic reptiles, surrounded in sepals, just like other scaled 

reptiles (Progga et al., 2021). They balance the ecosystems since they act as natural pest 

controllers. They are generally categorized as non-venomous and venomous. Non-

venomous have no venom or their venom has no adverse effects on humans. Venomous 

ones produce venom that they use mainly to subdue or kill prey rather than for self-defense 

(Abdurrazaq et al., 2019).  

Rajabizadeh and Rezghi, (2021) indicated that, out of 3848 known species of snakes, 

around 800 species are venomous, among which only about 50 species are deadly or fatal 

to humans. In Kenya, out of the 140 known snake species, 29 are considered venomous. 

Of the 29 venomous species, only 13 are of medical importance as their venoms cause 

injury or death in extreme cases, and only 9 of these account for bites that require medical 

attention (MOH-Kenya, 2019). 

Snakebite envenoming is a medical emergency that results from contact with snake 

venoms, which are injected when snake bites or when venom is sprayed into the victim's 

eyes (WHO, 2019). Krishnan, (2020) mentioned that snakebite envenoming is the second 

most deadly neglected tropical disease. In 2017, World Health Organisation (WHO) 

recognized snakebite envenoming as a Neglected Tropical Disease (NTD) that kills about 

1,410 to 137,880 people annually (Abdurrazaq et al., 2019). Furthermore, research 

indicates that about 5.4 million people suffer from snake bites annually, causing 1.8 to 2.7 

million cases of envenoming (Vasmatkar et al., 2020). 

In Kenya alone, 15,000 bites are approximated annually (MOH-Kenya, 2019). Progga et 

al., (2021) indicated that the deaths are mainly attributed to a lack of expert knowledge by 

the victims, on which type of snake one has encountered. Consequently, wrong 

identification of snake species can lead to wrong antidote administration which can in turn 

cause an adverse medical reaction. Therefore, easy and correct identification of the biting 



 
 

2 

 

snake species is a crucial part of the snakebite envenoming treatment process. When the 

snake in question is correctly classified based on the type of snake or the venom it 

contains, medical practitioners have an easy time administering the suitable anti-venom 

and carrying out correct first aid (Abdurrazaq et al., 2019). 

However, Rajabizadeh and Rezghi, (2021) noted that it is difficult to classify snake 

species unless one is guided by expert knowledge. Therefore, there is a need to apply 

machine learning techniques to make use of this expert knowledge to help the public 

without access to this knowledge classify snakes according to their species and based on 

the venom type that they contain. With such a model, snake species classification is 

automated, knowledge of the venom category present in a snake species is made instant, 

the treatment process becomes efficient, and ultimately more lives are saved. 

In the recent past, various models have been built to aid in snake species classification. 

These models made use of image processing techniques to come up with a classification. 

The images are captured from a camera device and run through algorithms to classify the 

snake (Joseph, 2021). The use of raw data as direct input to a classifier yielded low 

accuracy because of the varied features present on the snakes even of the same species 

(Rajabizadeh & Rezghi, 2021). Progga et al., (2021) proposed the use of feature extraction 

algorithms together with deep neural networks like CNN as a solution to this. 

Nath et al., (2014) mentioned that feature extraction is an important step in image 

classification. It is the identification, extraction, and analysis of the most relevant image 

features of an image and assigning it to a label (Ahmed Medjahed, 2015). Generally, 

snakes have varied dorsal color patterns, elongated and flexible bodies, and usually pose 

in varied ways causing deformation of the body making acquiring features quite 

challenging. To overcome the challenge of wide variations in the body pose of snakes in 

the images, a feature extraction method was used to extract features from images for 

indexing and retrieval which was later used for classification (Rajabizadeh & Rezghi, 

2021). 

Dimensionality reduction approaches such as Principal Component Analysis (PCA) and 

Linear Discriminant Analysis (LDA)] were employed by Rajabizadeh and Rezghi, (2021) 



 
 

3 

 

as the feature reduction technique in conjunction with transfer learning such as  

MobileNetV2 which yielded an accuracy of 93.16%. On the other hand, Mateen et al., 

(2019) used Singular Value Decomposition (SVD) in image classification as a feature 

extractor to decrease dimensionality based on reliability. The features that were not 

reliable for the classification were dropped and consequently the dimension of the image 

was reduced making classification easy and faster. Despite this apparent advantage present 

on SVD, it has not yet been exploited in studies classifying snakes based on their images. 

It is on the same premise that this research is built on. The research used SVD to reduce 

the feature of the snake on an image. The acquired vectors were fed as an input to train 

our deep neural network using MobileNetV2 transfer learning. The main advantage of 

SVD is, the ability to rapidly reduced the number of parameters as well as reduced the 

number of computations in complex data like that of snake image (Mateen et al., 2019). 

Moreover, since the snake image dataset was relatively small, MobileNetV2 transfer 

learning was considered because of its ability to train a deep neural network with little 

data by applying weights of another model that achieved better results. Snake features are 

not easy to distinguish and describe hence difficulties in classifying them accurately and 

consequently administration of treatment (Abdurrazaq et al., 2019) (Abayaratne et al., 

2019). 

1.2 Problem Statement 

Snake species are best classified based on their venom toxicity profile. There are five 

broad venoms toxicity profile categories. These include neurotoxic, hemotoxic, cytotoxic, 

cardio-toxic, and non-toxic. Neurotoxic destroys the nerves, hemotoxic causes abnormal 

blood changes, cytotoxic affects the muscle tissue, cardio-toxic inhibits the normal 

functioning of the heart, and lastly nontoxic are the ones that don’t have toxins that affect 

human beings (MOH-Kenya, 2019). Administration of the antidote and first aid measures 

are different for each envenomation (Abdurrazaq et al., 2019). Correct identification of 

the snake species is a crucial step for correct and proper first aid and treatment of snake 

bites (Abayaratne et al., 2019). 

Accurate classification of snake species is a challenging task for both humans and 

machines mainly because of subtle differences between different snake species and strong 



 
 

4 

 

variety within the same species. Some distinguishing features are so subtle to distinguish 

in a quick flash or poor lighting and may get deformed or simply have wear and tear. This 

makes it difficult to make correct identification in the field. As such, only a detailed 

examination of the captured specimen image can assist in making a positive identification 

(Rajabizadeh & Rezghi, 2021).  

Image classification, powered by advances in deep neural networks, is now performing 

many vision-related tasks faster and better than humans mainly due to convolutional 

neural networks (Devikar, 2016). Currently, dimensionality reduction methods are also 

used to perform feature reduction on images with wide varied features. PCA and LDA 

coupled with transfer learning of MobileNetV2 yielded an accuracy of 93.3%. SVD is 

however not exploited despite its apparent advantage of reducing the number of 

parameters besides reducing the number of computations in the complex image data while 

maintaining the reliability of the compressed image (Mateen et al., 2019). SVD is based 

on low-rank matrix completion using singular value truncation and thresholding. Since 

the image is decomposed using singular value decomposition (SVD) to obtain a low rank 

of image data, which is approximated in compressed form which is later on exploited to 

retrieve visual quality of the compressed image using singular value thresholding 

algorithm the reliability of the compressed image is not compromised. In machine 

learning, having abundant and correct data is more important than having correct 

algorithms. However, in instances where the dataset is relatively small, then transfer 

learning is applied. As such, this research focused on solving the classification of snake 

species by first creating an image dataset of Kenyan snake species. The features of the 

image were then reduced using Singular Value Decomposition (SVD) and saved to a 

matrix. The matrix with the reduced features of the images was then used to train a 

convolutional neural network via supervised transfer learning using MobileNetV2. 

1.3 Objectives 

1.3.1 General Objective 

This research aims to build a snake species classification model for snakebite envenoming 

management. 



 
 

5 

 

1.3.2 Specific Objectives 

(i) To investigate challenges associated with managing snakebite envenomation.  

(ii) To analyze machine learning techniques that support image classification. 

(iii) To develop an image classification model that can classify snake species. 

(iv) To validate the image classification model. 

(v) To develop a mobile application to show case the working of the model. 

1.4 Research Questions 

(i) What are the challenges with snakebite envenomation management?  

(ii) What are the machine learning techniques that support image classification?  

(iii) How can an image classification model that identifies snake species be 

developed? 

(iv) How can an image classification model be validated? 

(v) How can the mobile application be developed? 

1.5 Justification 

Firstly, the model from the research can assist the victims of snakebite envenoming to 

correctly identify and classify the snake encountered and seek appropriate first aid.  

Secondly, the public can benefit from this research by being able to correctly classify the 

snake species using the mobile application and administer the correct first aid to the 

victims. 

Thirdly, health workers can benefit from the research by being able to know the snake 

species the patient has encountered and the type of venom it contains and therefore 

administer the correct antidote based on the image collected by the patient. This can 

enhance service delivery in hospitals and consequently reduce deaths and disability caused 

by delayed envenoming management.  

Lastly, the study has contributed to a Kenyan dataset that can be used by academicians in 

the future for improving the model and carrying out research using the dataset. The finding 

of this study can be of importance to academicians and future researchers by contributing 

to the local and international body of literature. It can add value to the existing literature 



 
 

6 

 

by providing new techniques based on applying ICTs in snakebite envenomation 

management. 

1.6 Scope and Limitation 

The research is limited to classifying snake species using captured image and provides the 

common name of the snake in English and their scientific name and indicating their venom 

toxicity category. The research will further use a subset of five venom toxicity categories 

of snake species, which includes the eighteen endemic snake species in Kenya, for the 

proof of concept. For one to use the system, one must have a devices that has camera 

capability or a smart phone with clear camera. Also, there is need for initial download and 

installation of the application that uses the model. 

 

 

 

 

Chapter 2: Literature Review 

2.1 Introduction 

This chapter briefly introduced snakebite envenomation management, the challenges 

associated with it, and how technology has and can play a role in snakebite envenomation 

management. A review of the image classification and common methods currently in use 

was done. A comparison of different machine learning techniques that support image 

classification was conducted. This helped the researcher identify the suitable image 

classification model for snake classification. Finally, the researcher presented a conceptual 

framework.  

2.2 Snakebite Envenomation Overview 

Krishnan, (2020) mentioned that snakebite envenoming is the second most deadly 

neglected tropical disease. It is considered a medical emergency that results from contact 

with the venom of a venomous snake being injected either through a snake bite or by 



 
 

7 

 

venom being sprayed into the victim's eyes (WHO, 2019). Research indicates that about 

5.4 million people endure snakebites annually, causing 1.8 to 2.7 million cases of 

envenoming (Vasmatkar et al., 2020). In Kenya alone, 15,000 bites are approximated 

annually, and more go unreported (MOH-Kenya, 2019).  

Community awareness campaigns focusing on the prevention of snakebite such as cutting 

grass, and decluttering thickets have largely been done by most concerned organizations 

but still, the envenomation rate is on the rise, especially in low-income populations  (Pucca 

et al., 2020). Snakebite envenoming is largely reported by farmers, pastoralists, and 

fishermen among others. This is largely attributed to the fact that their normal duties 

revolve around the same habitat as the snake species. As a result, snakebite envenoming 

was considered an occupational disease  (Bhargava et al., 2020).  

Pucca et al., (2020) indicated that since most of the snakebite envenoming victims come 

from poor backgrounds, they normally shy away from seeking medical attention in health 

facilities near them. This accounts for the huge unrecorded number of snakebite 

envenoming. Instead, the victims seek attention from traditional healers whom they 

believe to be affordable and knowledgeable in matters of snakebite envenoming 

management (Bhargava et al., 2020). 

 Russell et al., (2021) indicated the importance of first aid management of the snakebite 

envenoming that occurs in the fields. He argued that correct and appropriate first aid done 

in the field contributes largely to the survival of the victim, especially when dealing with 

a fast-acting venom. First aid is usually done solely to keep the victim alive until they get 

medical attention. After correct administration of the first aid, the victim should be taken 

to a health facility for treatment or treated traditionally by herbalists or family members. 

The management of the snakebite envenoming is normally guided by correct identification 

and classification of the biting snake. 

2.2.1 Traditional methods of Identifying snake  

Bolon et al., (2020) mentioned that the most important step in snakebite envenoming 

management, whether in a health facility or by a traditional herbalist, is the correct 

identification of the snake and classifying it based on the right venom toxicity category. 

However, this is normally a challenging task for most victims because of a lack of expert 



 
 

8 

 

knowledge (Ahmed Medjahed, 2015). Different methods have been used by the victims 

to identify the biting snake. One of the common methods used in the identification of the 

snake species is killing and capturing the biting snake and taking it to a health facility 

alongside the victim. This is normally done by the victims to assist health workers to 

identify the correct anti-venom in the shortest time possible. However, WHO, (2019) 

snakebite management guidelines stipulate that it is not appropriate for anyone to handle 

snakes whether dead or alive not unless you are an expert in that area. This is because 

even a dead snake can cause envenomation. Aside from that if the snake is not completely 

dead, it can strike and cause more harm to the victim or people nearby. 

Another commonly used method in snake identification is a verbal description of the 

biting snake by either the victim or bystanders (Bolon et al., 2020). This involves 

describing the visual features of the biting snake e.g., color, size, pace, head shape, etc. 

However, victims are usually in a state of panic and confusion and might not describe the 

snake encountered well and exhaustively (Young, 2019). Aside from that, the description 

might be varied from one person to the other as much as they were all present at sight. 

Also, snakes from the same species might have varied visual features and of varied lengths 

depending on gender and age (Vasmatkar et al., 2020). 

Biting snake identification can also be achieved by observing signs and symptoms the 

patient portrays and associating them with a particular type of species and the toxicity of 

the venom they have (Russell et al., 2021). Different venom toxicity has different signs 

and symptoms (Muguti, 2003). Cardiotoxic inhibits the heart’s normal functioning, 

neurotoxic venom destroys the nerves, hemotoxic causes abnormal blood changes, 

cytotoxic affects the muscle tissue and nontoxic have little to no effect on human beings 

(Nag & Karforma, 2016).  The drawback of this method is, that victims might exhibit 

signs e.g., shortness of breath due to panic and fear and not necessarily because of the 

venom injected. This will in turn give a wrong identification of snakes using symptoms 

exhibited by the victim. 

Nag and Karforma, (2016) stated that testing the blood samples of the victim for the 

presence of venom antigens and associating it with the snakes that have certain venom 

toxicity found is another way of identifying and classifying the biting snake species. 



 
 

9 

 

Thereafter administration of the anti-venom is done based on the venom antigen found 

(Bolon et al., 2020). This is an accurate and more precise method. The only challenge with 

this method is that it can only be done in a health facility and requires laboratory 

equipment and expert knowledge. Aside from that, the cost associated with it can be 

punitive considering the persons affected by snakebite are generally of poor social-

economic status. Furthermore, the time it takes for laboratory results to be out might not 

be favourable for a fast-acting venom case hence minimizing the survival chance of the 

victim. 

Rajabizadeh and Rezghi, (2021) suggested that the use of pictures of the biting snake taken 

by the victim or bystanders can be used to classify the snake. The picture can then be given 

to the person with expert knowledge for correct analysis and classification. Once the 

herpetologist, correctly identifies and classifies the snake, a suggestion of the antidote to 

be administered is given. In the recent past, machine learning techniques have been used 

to classify snake species based on their image. The major challenge however is that the 

accuracy of the built models has been low due to the fact that snakes have varied features 

that are present even in the same species. 

2.2.2 Challenges of Snakebite Envenomation Management 

Snakebite envenomation management is the process of managing the venom injected into 

a victim by the snake (MOH-Kenya, 2019). This is normally done by the health officer, a 

traditional herbalist, family members, or any person present at the site. Bhargava et al., 

(2020) reinstate the importance of the first aid done at the site to increase the victim’s 

survival chance and minimize the severity of the disability caused by envenomation. The 

type of snake identified, and the venom toxicity it contains, dictates the type of first aid to 

be carried (Bolon et al., 2020). The majority of the public and the victims lack the expert 

knowledge to correctly classify the snake species and predict its venom toxicity 

(Rajabizadeh & Rezghi, 2021). As such, victims end up receiving the wrong first aid 

which affects the treatment process (Bhargava et al., 2020). Dissemination of expert 

knowledge easily and conveniently is needed to increase the survival chances of the 

victims. 



 
 

10 

 

Further, Rusli et al., (2019) mentioned that when handling snakebite envenoming, time 

becomes of the essence when dealing with snake species that have fast-acting venom 

(Muguti, 2003). First and correct treatment becomes critical to victims' survival. Also, 

considering most snakebite envenoming victims come from poor socioeconomic status, 

they normally shy away from getting treatment at the health facility and instead opt to try 

traditional ways of treatment (Bolon et al., 2020). Some of these methods end up making 

the wound area infected and can lead to permanent disability in the long run. 

Bhargava et al., ( 2020) and Ameade et al., (2021) indicated that most health workers have 

little knowledge about snakebite envenoming and the appropriate antidote to administer 

to patients in case of snake envenoming emergency. Bolon et al., (2020) indicated that 

wrong antidote administration can cause serve allergic reactions. As such, having correct 

knowledge about the snake biting species and its venom toxicity content becomes 

important to the treatment of the victim. Considering the challenges mentioned above, the 

use of technology in snakebite envenomation management has been researched 

exhaustively and models and methods have been proposed. 

2.3 Application of Technology in Snakebite Envenomation Management 

Technology has been applied in various ways to solve different challenges in different 

sectors. In snakebite envenomation, technology has been used to address the two most 

important steps to envenoming management i.e., classification of the biting snake species 

and the dissemination of the right information to aid in treatment and first aid. Some of 

the methods include using machine learning in natural language processing to classify 

snakes based on victims' descriptions and use of image classifiers to classify snake species 

based on their images. 

2.3.1Use Natural Language Processing (NLP) 

Rusli et al., ( 2019) developed an intelligent system that predicted the type of the snake 

by taking into account the description of the biting snake in the unstructured text, using 

natural language processing. Common perceptions and words used by several diverse 

groups of people to describe many different snake species were analyzed. The text was 

then pre-processed, relevant keywords extracted based on their weight in the context and 

later used as features during classification by machine learning to learn and predict the 



 
 

11 

 

species of snakes being described. The major drawback of this method is that it’s 

dependent on visual descriptions of the snake and some snake species have varied 

different features even within the same family based on their age, location, and sex.  

Niteesh. et al., (2021) on the other hand proposed a machine learning model that used the 

ID3 algorithm and the snakes’ dataset which was embedded into the model, and the model 

was trained with the snakes’ dataset. An input query to the trained model needed to include 

a description of the snake. The victim needed to mention the head shape, size of the snake, 

movement, pattern color on the skin, and color of the snake, then the model predicts the 

snake type so that the victim can take proper first aid treatment. A major drawback of this 

model is reliant on the ability of the victim to see and describe the feature of the biting 

snake correctly. This is usually a challenge since most victims are usually in a panic state 

after the snake bite and might be keen to correctly identify the biting snake or have a vivid 

memory of what they saw. 

2.3.2 Use of Image Classifier 

Image classification is a sub-group of computer vision that categorizes and labels groups 

of pixels or vectors within an image using a collection of predefined tags or categories 

and the visual features present. Image classification has been applied in several fields 

including plant disease detection, tumor identification, etc (Nyaga, 2019) (Nyaga, 2019). 

The image classification techniques are divided mainly into three categories. Supervised, 

unsupervised, and semi-supervised. Supervised classification is where model training 

includes showing the model the images along with their corresponding label. 

Unsupervised classification is where there is no labeling of the data points and the 

algorithm uses raw data for training. Semi-supervised classification considers the 

advantages of both supervised and unsupervised (Sanghvi et al., 2021). 

As manual extraction of features describing the appearance of a snake is tedious, recent 

articles used automated feature extraction. Bloch et al., (2020) classified 22 snake species 

that were manually collected from Perlis Snake Park in Malaysia. Features of these 

collected images were extracted using Color and Edge Directivity Descriptor (CEDD). A 

multiple supervised machine learning algorithm, such as Naive Bayes, k-Nearest 

Neighbors, etc. was then used as a classifier. However, in the traditional machine learning 



 
 

12 

 

technique, parameters in the feature extraction method are not trainable and need to be 

manually tuned.  

Abdurrazaq et al., (2019)  on the other hand classified snake images using a convolutional 

neural network (CNN). Three CNN architectures were evaluated using a dataset of 415 

snake images from five common hazardous venomous snake species in Indonesia. Five-

fold cross-validating shows that CNN is capable of classifying the snake images with high 

accuracy of 82%. The only drawback this method has is that CNN is a deep neural network 

model and a huge number of image datasets is needed to train the model. Aside from that, 

it is expensive both in the computational time it takes to train and test the model. 

Based on the above literature,  accurate identification of the features present in an image 

is the major objective of image classification (Nath et al., 2014).  The main principle of 

an image classifier is to recognize the feature occurring in an image with the help of 

different mathematical techniques and use them to classify the image accurately. The steps 

for implementing the image classification process are shown in Figure 2.1 below.  



 
 

13 

 

 

Figure 2. 1: Image Classification Process (Nath et al., 2014) 

2.4 Steps of Classifying Image 

2.4.1 Image Pre-Processing 

Nath et al., (2014) indicated that pre-processing of image data is an essential part of the 

image classification process. It is normally done to reduce complexity and increase the 

accuracy of the applied algorithm. Aside from this, it also allows better feature extraction 

during classification (Vasmatkar et al., 2020). Raw data is normally characterized by 

unwanted distortion and noise. It is pre-processed by reshaping, resizing, and augmenting 

the image. Image Augmentation is the technique of creating new images from already 

existing images for training a deep learning model (Progga et al., 2021). This is normally 

done to avoid overfitting the model and to increase the size of the dataset (Bhargava et al., 

2020). Due to the randomized transformations, the same images cannot be obtained each 



 
 

14 

 

time. Good model accuracy is dependent on the pre-processing of the raw images that are 

fed as input into the classifier (Progga et al., 2021). 

2.4.2 Feature Extraction. 

Ahmed Medjahed., ( 2015) defined feature extraction as the process used to reduce the 

amount of data needed to describe a large set of data correctly with an aim of facilitating 

decision-making, e.g., pattern classification. The most relevant feature of an image is 

usually extracted using mathematical functions and assigned to a label for the 

classification model. Bloch et al., (2020) did feature extraction of the snake images using 

Color and Edge Directivity Descriptor. Rajabizadeh and  Rezghi, (2021) explored the use 

of Principal Component Analysis and Linear Discriminant Analysis, to extract features of 

an image that were later fed as vectors to image classifiers. LDA performed well compared 

to PCA. Also, Singular Value Decomposition has been used to reduce the data redundancy 

inherent in the dataset. A good feature extraction technique will produce good relevant 

results (Farajzadeh et al., 2017). 

2.4.3 Image Classification 

The vectors acquired from the feature extraction process are usually fed into a classifier 

of choice and labels assigned to them. The images are then categorized into predefined 

classes using the selected classifier that compares the image patterns with the learned 

labelled patterns. 

2.5 Feature Extraction Techniques 

Different techniques can be applied for image feature extraction. The dimensionality 

reduction processes are important in reducing computational time and storage. They can 

also be used in the feature extraction of an image. Feature extraction is the act of mapping 

an image from the image space to the feature space. We review the three commonly used 

dimensionality reduction techniques used in image feature reduction. 

2.5.1 Principal Component Analysis (PCA) 

PCA is an unsupervised linear technique that uses an orthogonal transformation to project 

a set of variables into a lower dimension with maximum variance. In PCA a set of 



 
 

15 

 

variables that are possibly correlated, convert into a set of values that are not correlated 

variables, called principal components.   

To reduce each data, the PCA tries to find an orthogonal matrix so that the reduced data 

have the maximum variance (Rajabizadeh & Rezghi, 2021) 

∁=
𝟏

𝑵−𝟏
∑(𝑿𝒊 − �̅�)(𝑿𝒊 − �̅�)T     (2.1) 

Where 

 �̅� is the mean. 

N is the number of data. 

2.5.2 Linear Discriminant Analysis (LDA) 

Linear Discriminant Analysis (LDA) is a statistical analysis technique that can be applied 

to reduce the dimensionality of hyperspectral data cubes and extract features for 

classification. LDA is a supervised dimensionality reduction technique that makes use of 

labels alongside the features in the dataset. LDA also aims to reduce high dimensionality 

by extracting low dimensional features which have the most sensitive discriminant ability 

from high dimensional feature space. For each data x, it tries to find an orthogonal 

projector U by minimizing the within-class distance and maximizing the between-class 

distance of the projected data. In this way, LDA can maximize the separability of species 

in the snake image datasets. Features that explain the highest variance among species are 

selected by LDA. LDA ranks features based on how much variation among the species 

they account for.  

2.5.3 Singular Value Decomposition (SVD) 

SVD is a reliable and robust orthogonal matrix decomposition technique. SVD is widely 

used for image analysis to solve the problems related to least squares, pseudo inverse 

computation of the matrix, and also in multivariate experiments. In the area of machine 

learning, SVD-based methods are used to decrease dimensionality, metric learning, 

manifold learning, and collaborative filtering (Farajzadeh et al., 2017). Singular Value 

Decomposition (SVD) is one of the matrix factorization methods that allows reducing the 

size while keeping the characteristics of a matrix. It is important to reduce the size of the 



 
 

16 

 

feature vector to reduce the computational load for training the classifier. Given an m × m 

matrix A, the expression of its SVD is as shown in equation 2.2 below. 

𝑨 = 𝑼 ∑ 𝑽T       (2.2) 

where U is an m×m matrix, 

 V is an n×n matrix a 

 Σ is the singular value of matrix A which is an m×n non-negative real diagonal matrix. 

We use SVD on the feature matrix, to make a shorter feature vector that still holds the 

features’ characteristics. The reason for using SVD to make feature vectors is that it is 

relatively straightforward to apply and gives the singular values, which contain 

information about the features generated in descending order. Ordering the elements of 

the feature vector in descending order gives an advantage when using a training data set 

for a neural network. SVD is a better option to decrease the dimensionality of huge 

amounts of data by speeding up the computational process. Particularly, in image 

classification, SVD also performs better in the form of classification accuracy with 

optimal computational time (Seok et al., 2012). 

2.6 Machine Learning Techniques used in Image Classification 

2.6.1 Support Vector Machine 

SVM is a supervised learning binary classifier in which, known labels help indicate 

whether the SVM is classifying correctly or not. It can be used as a prediction tool to 

predict future data. SVM predicts based on machine learning theory that increases the 

accuracy of a classification and overfits to data automatically. SVM also performs 

mapping low dimensional space into high dimensional by using non-linear basis functions. 

SVM uses hypothesis space of linear functions on high dimension feature space that is 

trained with learning algorithm from optimization theory. SVM uses linear classifiers 

(hyperplanes) to separate the data (Mahajan & Chaudhary, 2019) 



 
 

17 

 

2.6.2 K-Nearest Neighbor (KNN) 

KNN is one of the most popular non-parametric, supervised machine learning algorithms 

in the field of pattern recognition, used for classification and regression. The classification 

rules in this algorithm are produced by only the training dataset. KNN predicts the test 

class according to the K training samples which are the closest neighbors to the test 

sample. The algorithm simply depends on the distance between feature vectors and 

classifies unknown data points by finding the most common class among the k-closest 

examples. The prediction is based on the largest class probability. The probability is given 

by the equation 2.3 below.  

𝑷(𝑿, 𝑪𝒋) =  ∑ 𝑺𝑰𝑴(𝑿, 𝒅𝒊). 𝒚(𝒅𝒊𝒅 , 𝑪𝒋)    (2.3) 

 

Where y(𝒅𝒊,𝑪𝒋) is a category attribute function 

The similarities are calculated between all training samples and X.ith 

sample 𝒅𝒊(𝒅𝒊𝟏, 𝒅𝒊𝟐, … … 𝒅𝒊𝒏) 

 

𝑨𝒏𝒅 𝑺𝑰𝑴 (𝑿, 𝒅𝒋) =
∑ 𝑿𝒋  𝒅𝒊𝒋

𝒎
𝒋=𝟏

(√(∑ 𝑿𝒋
𝒎
𝒋=𝟏 )𝟐 √(∑ 𝑿𝒅𝒋𝒊

𝒎
𝒋=𝟏 )𝟐)

       (2.4) 

 

Training categories C1, C2, C3,…,.Cj and sum of training samples is N 

Sample X is the same feature vector of the form (X1, X2, X3…., Xm) 

2.6.3 Convolution Neural Network (CNN) 

The Convolution Neural Network (CovNet) is a widely used multi-layer neural network, 

designed to recognize visual patterns directly from pixel images with minimal pre-

processing for implementing the deep learning technique. It comprises layers for feature 

extraction and classification. The advantage of CNN is its ability to identify the image 

features without supervision from the human which is an advantage to snake image 



 
 

18 

 

classification for envenomation management (Manoj et al., 2018). Figure 2.2 below shows 

the architecture of a CNN image classification model. 

 

Figure 2. 2: CNN model Architecture (Progga et al., 2021) 

A CNN consists of different layers, which include convolutional, an activation function 

usually rectified linear unit (ReLU), pooling, and the fully connected layers. 

 Convolution layer – This layer extracts features from the input image. Filters are used 

to extract these features. For example, assume a 32 x 32 image. The computer sees a 32 x 

32 x 3 array of numbers, where the 3 represent the red, green and blue values for each 

picture element (pixel). Assume a 5 x 5 x 3 filter. The filter takes the first position at the 

top left corner of the image. It then slides, or convolves, across the input image multiplying 

the values in the filter with the original pixel values of the image. These multiplications 

are summed up to create one figure that represents where the filter is. The filter moves 

one unit or stride to the left and the process repeats. After sliding over the whole image, a 

28 x 28 x 1 array of numbers called the feature map is created. If 6 5 x 5 filters are used 

then a 28 x 28 x 6 array will be created (Nielsen, 2015). 

After each convolution, the output reduces in size and becomes the input of the next layer.  

Activation function – Rectified Linear Unit (ReLU) – ReLU is an operation that replaces 

all negative pixel values in the feature map by zero. This is to introduce non-linearity since 

most of the real-world data is non-linear (Devikar, 2016). Pooling Layer – Pooling layers 

usually come after the convolution layer and its role is to simplify the information in the 

output from the convolutional layer. It reduces the dimensionality of each feature map but 



 
 

19 

 

retains the most important information. A common pooling procedure is max pooling, 

where a pooling unit outputs the maximum value (Nielsen, 2015). 

Pooling Layer-This layer comes after every convolutional layer. The role of the pooling 

layers is to simplify information in the output from the convolutional layer. The 

simplification process is achieved by the pooling layer taking each feature map output 

from the convolutional layer and preparing a condensed feature map. For instance, more 

precisely, each unit in the pooling layer may summarize a region of neurons in the 

previous layer. Max-pooling is one of common procedures used in the pooling layer. Max-

pooling involves a pooling unit simply outputting the maximum activation in the 2 × 2 

input region (Nielsen, 2015). For instance, considering 24 × 24 neurons output from the 

convolutional layer, after pooling, this will result in 12 × 12 neurons. This process reduces 

the dimensionality of each feature map while at the same time retaining important 

information. 

Fully connected Layer – After several convolutional and pooling layers a fully connected 

layer takes all the neurons in the previous layer and connects it to every neuron in the layer 

to produce fully connected output. This is the layer that does the actual classification work. 

Kamalraj, (2020) classified snake species using snake bite marks through image 

processing and Convolution Neural network. A picture of snake bite marks was uploaded 

to the system and later underwent image pre-processing. A convolution neural network 

was then used to classify the snakes as either venomous or non-venomous. However, 

getting a clear image of a snake bite is difficult because most patients tamper with the bite 

site while doing first aid. Furthermore, rarely will patients or doctors take images of the 

bitten site. Therefore, its biggest drawback becomes the difficulty in the collection of the 

images of all snake bite marks on individuals making the dataset too small to warrant 

model training. Also, the collection process has to be accurate and requires high-quality 

images for the process (Niteesh. et al., 2021). 

Vasmatkar et al., (2020) proposed a model that involved techniques based on Image 

Processing, Convolution Neural Networks, and Deep Learning to classify snakes. CNN 

was used to extract features and utilize them for classification. However, deep learning 



 
 

20 

 

methods like CNN require one to have a large dataset for better accuracy. This is a 

challenging task considering the risk and difficulty of collecting snake image data. 

2.6.4 Transfer Learning 

The drawback of implementing CNN on a small dataset is that it results in overfitting 

(Bhargava et al., 2020). A possible way of overcoming this challenge is to increase the 

size of the training data. However, this is a difficult and expensive task (Kaur & Gandhi, 

2020). As such, Transfer learning becomes the appropriate solution. Zhuang et al., (2021) 

defined transfer learning as the use of stored knowledge gained from a pre-trained neural 

network model, to train a different but related neural network model with a smaller dataset. 

A major advantage of using Transfer learning is that it reduces the time a neural network 

takes to train a model and also results in a lower generalization error (Krishna & Kalluri, 

2019). Sharing of pre-trained models is normally achieved with the help of millions of 

parameters or weights (Kaur & Gandhi, 2020). Figure 2.3 below shows a representation 

of Transfer learning.  

 

Figure 2. 3 : Transfer Learning Architecture (Ribani & Marengoni, 2019) 

There are different models for transfer learning. Among them, DenseNet, MobileNetV2, 

VGG19 etc. DenseNet is a convolutional neural network model made up of dense blocks 

with a densely connected layer that uses a smaller number of parameters. Every dense 

block has a batch normalization layer, rectified linear unit (ReLu) activation, and 3x3 



 
 

21 

 

convolution forming a composite function. In DenseNet, all the feature maps are 

concatenated, but the feature maps of each layer have the same size. Transition layers 

between two blocks contain batch normalization, 1x1 convolution followed by average 

pooling ensures downsampling. Bottleneck layers have been introduced before each 3x3 

convolution to reduce the number of input feature maps and thus, improve computational 

efficiency. Vasmatkar et al., (2020) classified snakes images using CNN transfer learning 

DenseNet161 but applied the GrabCut algorithm to the images as a feature extractor and 

performed data augmentation. 

 MobileNetV2 is a light weighted model mainly used for mobile and embedded vision 

applications. The model is built on depthwise separable convolution. It introduces two 

global hyperparameters, which are width multiplier and resolution multiplier which helps 

to trade-off between latency and accuracy. Rajabizadeh and Rezghi, (2021) explored the 

combination of convolutional neural networks that was achieved through transfer learning 

using MobileNetV2 and dimensionality reduction approaches for feature extraction of the 

images. The model achieved an accuracy of 93.16%.  

2.6.5 Tensor Flow Framework  

TensorFlow is a high-performance numerical computation framework or software library 

originally developed by researchers and engineers at Google. It is an end-to-end platform 

that is open source and has strong support for the deployment of machine learning and 

deep learning models. It offers a free GPU which makes it faster and easier to train a 

machine learning model compared to a CPU (TensorFlow, 2018). At Google, for example, 

the framework powers many of their products like Gmail and Google translate. This study 

used the convolutional neural network libraries and tensorboard, the visualization toolkit, 

of TensorFlow. 

2.6.6 Tensor Flow Lite. 

Tensor Flow lite is a set of tools that facilitates on-device learning, by enabling models to 

run on the mobile, embedded or edge devices. They are generally fast because the speed 

and storage are optimized. It provides the basic features needed for making inferences at 

the edge like, low latency, lightweight, optimal power consumption, secure among others. 

After the model is trained, it is converted to a tensor Tensor Flow Lite format using a 



 
 

22 

 

converter. During the conversion process from Tensor Flow model to Tensor Flow Lite 

model, the side of the file of the model is reduced. There is normally a choice of the trade-

off between reduced file size and execution speed. Tensor Flow Lite model is efficient in 

accuracy and is also a lightweight version that occupies less space hence its popularity in 

use for mobile devices.(Choe et al., 2022; David et al., 2021)  

Table 2. 1 Summary of Literature Reviewed 

Author Technique Used Weakness 

Rusli et al., ( 2019 NLP Visual descriptions of 

snake vary depending on 

environment and age, sex 

etc. 

Bloch et al., (2020) CEED &Naïve Bayes Traditional machine 

learning parameters in 

feature reduction method 

are not trainable 

Abdurrazaq et al., (2019) CNN Needs a huge dataset 

Rajabizadeh and  Rezghi, 

(2021) 

PCA&MobileNetv2 Low accuracy 

Niteesh. et al., (2021) 1D3 Victims give incorrect 

descriptions when in panic 

mode 

2.7 Research Gap 

Classification of snakes is still a major challenge without expert knowledge and relying 

on the human eye and descriptions has proven to be a daunting and difficult task. One 

would require expert knowledge to achieve this. The deep learning algorithms discussed 

in the preceding sections describe recent research on the classification of snake species 

through the use of images. However, the availability of snake specialized datasets is a 

challenge more so regarding the rare snakes. Moreover, extracting the distinct feature of 

snake species is very important to come up with better inspection methods. Previous 

research has explored the use of PCA and LDA but didn’t explore the possibility of using 



 
 

23 

 

SVD despite its apparent advantages. Considering the small dataset at hand, this research 

proposed to explore feature extraction using SVD and later feed the vectors as input to a 

deep neural network using transfer learning to classify snakes according to their name and 

venom toxicity category, based on their images. This is achieved by harnessing the power 

of transfer machine learning. 

2.8 Conceptual Framework 

The following conceptual framework links the reviewed literature with the research 

problem and objectives. Based on the literature review, the framework shown in figure 

2.4 combines techniques cited to come up with a model that can achieve the objectives of 

classifying snakes according to their species name based on the image. This framework 

was established to provide a solution to the problem in the study of classifying the snake 

species and venom toxicity based on their images. 

 



 
 

24 

 

 

Figure 2. 4 Conceptual Framework of the Snake Image classification Model 

 

 

  



 
 

25 

 

Chapter 3: Research Methodology 

3.1 Introduction 

This chapter explored the research design, data collection, and the development of the 

image classification model. Additionally, the chapter covered the methodology that would 

be used in system development. Finally, the chapter made a statement on the research 

quality and the ethical considerations for the research. 

3.2 Research Design 

A research design is a blueprint describing how a research study is to be completed: 

operationalizing variables so they can be measured, selecting a sample of interest to study, 

collecting data to be used, and analyzing the results (Thyer, 1993). The objective of this 

research is to design an image classification model that can classify snake species for 

snakebite envenomation management. This is applied research because it solves the real-

life problem of managing snakebite envenoming. The research paradigm is quantitative in 

nature and the data collected from the images taken can be quantified in numbers to 

represent the output. The method is correlational in nature because features from the snake 

images can be used to classify the outcome of an image. 

3.3 Population and Sampling 

3.3.1. Population 

Several but quite limited studies have been done on snake species worldwide and none so 

far to the best of the researcher’s knowledge had been conducted in Kenya. Considering 

the difficulties and risks experienced in capturing the image of these stealth and dangerous 

species, the dataset is normally small and not varied. Furthermore, there is currently no 

secondary dataset that exists of images of Kenyan snake species to the best of researcher’s 

knowledge. Therefore, primary data was used for the study. The snake species images will 

be the target population mainly from the Kenya National Museum of Kenya and the 

Stedmark snake park Garden in Karen. The population consists of a total of 4521 images 

from eighteen different species that covered the five categories of snake venom toxicity. 

African Rock Python had 145 images, Black Mamba had 377 images, Black-necked 

spitting cobra had 128 images, Boomslang had 152 images, Brown House snake had 211 



 
 

26 

 

images, East African garter snake had 127 images,  Egyptian Cobra had 232 images,  

Gabon viper had 253 images, Green Mamba had 428 images, Jackson Tree snake had 57 

images, Jameson Mamba had 381 images, Large Brown spitting cobra had 433 images, 

Mt. Kenya bush viper had 203 images, Puff Adder had 187 images, Red spitting cobra 

had 331 images, Rufous beaked snake had 396 images, Sand Boa had 160 images, and 

finally western cobra had 320 images.  

3.3.2 Sampling 

The sampling technique used was a combination of clustered and convenient sampling of 

the data. This is because non-probability sampling supports its use where the researcher 

has limited control of the representation of the selection. Snakes generally hide and reside 

in forests and thickets hence the difficulty in getting the images. Furthermore, they are 

dangerous and wild and can attack when they feel threatened. Therefore, the snake species 

available in the National Museum and Stedmark Garden snake park in Karen became a 

representation of the snake species found in Kenya. The species available in these two 

snake parks have been picked from different parts of Kenya including but not limited to, 

Meru, Nairobi, Kibwezi, Mombasa, Kakamega, etc. The primary images were collected 

under the guidance of a herpetologist and with the help of a snake handler for proof of 

concept.  

3.4 Data Collection 

Primary data consisting of the snake images, venom toxicity category, common name, 

type of general antidote administered, first aid guidelines, and scientific names were 

collected from Kenya National Museum, Nairobi, and Stedmark Gardens snake Park, 

Nairobi. The primary devices for data collection from the sources were digital cameras 

and smartphone devices which is the target device for the application. The image files 

collected were in .jpg and .arw format since the quality of the images was of high 

importance to building a good model. A sample of the primary data collected is attached 

in appendix G. The image of the snake species type was saved in a folder using its common 

name. An excel file was created and it contained the snake species' common name, 

scientific name, and the venom toxicity category of the species. Since the users are not 



 
 

27 

 

experts, the model predicts the snake species on the image and classifies it based on the 

venom toxicity category on their behalf.  

3.5 Image Pre-Processing and Analysis 

3.5.1 Image Augmentation 

Image augmentation artificially expands an image dataset. Image augmentation operations 

include flipping, rotation, shear, zoom, cropping, deforming, adjusting the hue, adjusting 

saturation, or adjusting brightness and contrast (Custom Image Augmentation: Towards 

Data Science, 2017). Considering the small dataset available for the research, the images 

were augmented to expand the image dataset by three or four times and helped the model 

cope with all distortions that occur in real-life images. 

3.5.2 Standardization  

The images were cropped to remove excess background and ensured at least 60% of the 

image coverage size was composed of the snake. The raw images were of different formats 

and therefore were converted to a standard format .jpeg also which is the format of our 

target device. The images were pre-processed as well to increase the brightness and 

contrast and resized to 224*224, which is the image size input requirement for 

MobileNetV2. Feature reduction was implemented using SVD, to pick the most relevant 

feature for the classification. These images served as an input to the model for 

classification.  

3.5.3 Folder Structure 

Approximately 70% of the images formed the training dataset and the remaining 30% 

formed the validation dataset. The images collected were put into the respective dataset 

folder and subfolders corresponding to the species name.  

3.6 System Development Methodology 

3.6.1 Agile System Development  

An agile development methodology was used in this research because it is incremental 

and iterative. Most importantly it has a theoretical background that is consistent with the 

problem-solving approaches tackled in this research. 



 
 

28 

 

3.6.2 System Analysis  

The system detected the features on the snake images and classified the snakes based on 

their name and venom toxicity. The camera was the primary device used to capture the 

images needed for analysis. Therefore, the application classified a particular snake based 

on its species name and venom toxicity. 

3.6.3 System Design 

 The system architecture is comprised of both hardware components and software 

components. The hardware components are a camera for capturing images and a processor 

whereas the software components are the snake species classification model, the software 

tools used such as Google Colab, programming language, and its corresponding libraries.  

The snakebite envenomation management system which will be a mobile application, 

software modeling diagrams will be used to show the structure and logical interactions of 

the various components. System design modeling will be done using the following UML 

modeling diagrams. 

i. Use Case Diagrams: To provide a simplified graphical representation of what the 

system should do in a use case, which is a description of the tasks that a user can 

perform on the platform and how  

ii. Sequence Diagram: To show the interaction logic between the objects in the 

platform in the time order that the interactions take place 

iii. Entity-relationship Diagrams: To describe how various entities in the platform 

relate to each other 

iv. Class Diagrams: To model the objects that make up the system, display the 

relationships between the objects, and describe what those objects do and the 

services that they provide. 

3.6.4 System Implementation  

3.6.4.1 Model  

The development of the model was carried out on the Google Colab platform. Google 

Colab is a free cloud service that offers free GPU (Graphics Processing Unit) that were 

essential for building this model. The main input for the mobile application is the user-



 
 

29 

 

uploaded image, which is fed into the MobileNetV2 transfer model for classification. The 

result obtained is shown to the user. The application of deep learning required very 

expensive hardware and powerful libraries such as Keras, OpenCV, sklearn, and 

TensorFlow. This was created using the Google Colab and was the testbed environment. 

All training and validations were done in the same environment. 

3.6.4.2 Test Bed  

The testbed was required to test the model. This can include hardware and software. The 

hardware included a mobile phone with a camera. Software components included image 

classification algorithms were used to predict and classify different snake species and their 

toxicity category. The testbed was Google Colab where all test images and respective 

labels were inputted into the model to predict their classes. The testing was done using a 

test dataset that was set aside earlier during the research. Accuracy, precision, and recall 

were the metrics used to evaluate the performance of the model. 

3.6.4.3 Programming Language  

The programming language used in this research was python because of its wide array of 

open libraries that were easily accessed. This language easily works through the Keras 

and TensorFlow libraries. The algorithm can run on the Google Colab which offered 

graphics processing units for running the decay algorithm.  

3.7 Research Quality 

This was a very important aspect of maintaining accuracy in this research, therefore, 

validity, objectivity, and reliability of the research were important. This research achieved 

this by limiting bias in the data collected. The research minimized the effects of extraneous 

independent variables that were undesirable. This was achieved by using deep learning 

techniques that focussed on every aspect of the features of the images and the training can 

minimize their effects. Accuracy, precision, and loss function will be used to measure the 

validity of the model. A confusion matrix was used to generate a classification report in 

this research. 



 
 

30 

 

3.7.1 Research Validity  

Validity refers to the extent to which the data accurately measures what it was intended 

to. Automated model testing and a confusion matrix was used to ensure the validity of the 

model. Additionally, the MobileNetV2 network model’s loss or learning rate was visually 

tracked using Tensorboard; this made it easier to debug and optimize the model.  

3.7.2 Reliability 

Reliability focuses on the extent to which the data collection method will yield the same 

findings if replicated by others. Proper documentation of system requirements was done 

to ensure the reliability of the study. 

3.8 Ethical Considerations 

 Ethics are standards or codes of conduct that help distinguish between what is acceptable 

and what is not (Ethical Considerations: CIRT, 2018). Images used in this study were 

obtained from the National Museum of Kenya with permission to use for educational 

research. All previous works were cited appropriately and due acknowledgment was given 

to the respective authors. To ensure adherence to ethical codes of conduct, the collection 

process was in line with the permission for educational research. An approval from the 

Strathmore University Ethics Review Board attached in appendix E was obtained before 

commencing the study. To ensure the originality of the study, a similarity check was done 

and the report is as shown in appendix F. 

 

 

 

 

 

 

 



 
 

31 

 

Chapter 4: System Analysis, Design, and Architecture 

4.1 Introduction 

The main purpose of this study was to design a model that will classify Kenyan snake 

species using their image, based on their venom toxicity for the purposes of managing 

snakebite envenoming. Object oriented analysis and design (OOAD) research design was 

employed in this research. The chapter outlined the design architecture and explained the 

components of the developed system, the interactions between the different components 

of the developed system, and the interactions between the users and the developed system. 

Unified modeling language (UML) was employed to model and illustrate the interactions 

of the system using the use case diagram, system sequence diagram, entity-relationship 

diagram, and class diagram. Wireframes were also drawn to display the functional 

elements of the mobile application. 

4.2 Requirements Gathering and Analysis 

The primary raw images data were collected using a camera and a smartphone from the 

National Museum of Kenya and Stedmark Garden Snake Park, Kenya. The images had to 

capture the head and part of the body of the snake. Due to subtle differences that exist in 

snakes of the same species and similarities of different species, one way of distinguish 

between the snakes is to have a view of the body and the head together in image capture 

for better identification. They were in .jpeg and .arw format. They were all converted to 

standard format .jpg format using Adobe lightroom. Blurred, small, and unusable images 

were removed from the dataset. Images were manually cropped to get the best view with 

less background noise. They were categorized in folders and saved according to the snake 

species' local name. An excel file was created to document the snake’s common name, 

scientific name, and the venom present in that particular species. The images collected 

were of different shapes and therefore they were all reshaped to size (224x224x3) to fit 

the model input requirements. SVD was used to do feature reduction.  

4.2.1 Functional Requirements 

i. The system should allow first-time users to register.  

ii. The system should allow the user to upload a snake image in .jpeg or.jpg format 

as the input to the classification model. 



 
 

32 

 

iii. The system would predict the snake species in the input image. 

iv. The system should classify snake species according to their venom toxicity 

category. 

4.2.2Non-Functional Requirements 

i. Usability - The proposed users of the application are the general public and 

medical practitioners who would like to use it for snakebite envenoming 

management application. Therefore, the system should be easy to use. 

ii. Supportability Requirements - The model should support images from a wide 

range of input cameras with different formats. This will promote the usability of 

the model in a real-world environment.  

iii. Reliability Requirements - The model should be able to provide consistent results 

when provided with the same input data.  

iv. Maintainability The model should be easy to maintain and support.  

v. Scalability - The system should gradually improve on the accuracy levels. It should 

also adapt to new datasets from different environmental conditions. This can 

ensure a good application response to new data. 

4.3 System Design 

The MITRE Corporation, (2014) defined system design as visualizing the internal system 

or creating a blueprint of the system while deriving a solution that can be traced back to 

the user requirements. This research adopted agile development methodology which is 

incremental and iterative and therefore object-oriented analysis and design were used. 

4.3.1 System Architecture 

The system architecture for the snake image classification model is shown below in Figure 

4.1. Raw images from the camera or mobile phones are pre-processed and later split to 

create training and testing datasets. Image augmentation was done to increase the size of 

the dataset. Feature extraction was done using SVD and vectors fed to the model. 

Subsequently, the model was trained and validated using the transfer learning model 

MobileNetV2. The main input for the model is the user-uploaded images, which were fed 

into the model for classification. The result obtained was shown to the user.  



 
 

33 

 

 

Figure 4. 1System Architecture 

4.3.2 Use-Case Diagram and Descriptions 

A use case is a list of actions and event steps that shows and defines the interactions 

between an actor and a system to achieve a goal. The actors in this system are the general 

users and the developer. The developer collates and pre-processes the image dataset and 

trains the transferred deep neural network as explained in chapter three. The developer 

then creates a model that is used to infer the images uploaded by the user. This model 

makes the snake image classification. The victim, who is a sub-entity of the user, provides 

the system with an image to classify. Figure 4.2 below shows the system use case diagram. 



 
 

34 

 

 

Figure 4. 2Use case Diagram 

 

Table 4. 1 Register and Login Description 



 
 

35 

 

Use Case Register and Login 

Primary The User (Patient and Doctor) 

Brief Description This use case describes how the users will register and 

login. 

Pre-Condition The user wants to classify a snake image and know its 

details and antidote, Registration Details 

Post-Condition Registered user, Authenticated User 

Major Steps Performed 

Actor System 

The user enters 

Registration Details 

 

 System Verifies and saves registration details 

The user enters login 

details 

 

 The system verifies and authenticates the user 

 

Table 4. 2Load an Image and Classify Description 

Use Case Load an Image and Classify 

Primary The User  

Brief Description This use case describes how the users 

load an image into the system for 

classification. 

Pre-Condition Available Image, Trained model 

Post-Condition Classified Image with its information 

 Major Steps Performed 

Actor System 

The user selects an image from the photo 

file or takes an image using the phone's 

camera. 

 

 The system loads the image as input to 

the MobileNetv2 trained model 

User uploads a .jpg image for 

classification 

The system performs classification 

 The system displays the results 

The user views the result  

 

 

 

 

 

Table 4. 3 Query and View Results Description 



 
 

36 

 

Use Case Query and View Results 

Primary The User (Patient and Doctor) 

Brief Description This use case describes how the users 

query a snake image for classification and 

view the results 

Pre-Condition Logged in. Uploaded Image 

Post-Condition Image classified and results displayed 

 Major Steps Performed 

Actor System 

The user enters login details  

 The system verifies and authenticates the 

user 

User uploads a .jpg image for 

classification 

 

 The system queries the model and 

classifies the image 

 The system displays the results of the 

uploaded image and its information. 

 

Table 4. 4 Recommend Antidote Description 

Use Case Recommend Antidote 

Primary The Doctor 

Brief Description This use case describes how the Doctor 

will recommend the antidote based on the 

results of the queried image. 

Pre-Condition Result of the query image 

Post-Condition Recommended Anti-dote 

 Major Steps Performed 

Actor System 

 The system displays the results of the 

classified image with its venom toxicity 

information 

 The Doctor views the result of the 

classified image to identify which toxins 

it contains. 

 

The doctor recommends the appropriate 

antidote and logs it into the system 

 

 The system displays the antidote 

recommended  

User (patient) views the Recommended 

antidote 

 

 



 
 

37 

 

4.3.3 System Sequence Diagram 

The system sequence diagram shows the interactions between the main entities in the 

system. Figure 4.3 below shows the flow of activities in sequence on how the user gets 

registered and logs in to the system. 

 

Figure 4. 3 Login System Sequence Diagram 

 

 

 

 



 
 

38 

 

The patient selects or takes a picture that they want to be classified. The patient then 

loads the image to the image classifier model to classify the uploaded image. The image 

classifier queries the trained transferred neural network to obtain a result. The result is 

displayed to the patient. The doctor requests to view the result of the classified image 

and recommends an antidote based on the venom content of the classified snake. The 

patient gets to view the recommended antidote. This is shown in figure 4.4 below 

 

Figure 4. 4 Classifying System Sequence Diagram 

4.3.4 Entity Relationship Diagram 

Figure 4.5 describes information using a wide range of entities, attributes, and 

relationships. Entity Relationships (ER) are high-level conceptual models designed to 

facilitate the development of databases. The ERD diagram below, therefore, forms the 

blueprint of the database used in the detection model. 



 
 

39 

 

 

Figure 4. 5 Entity Relationship Diagram 

4.3.5 Class Diagram 

A class diagram defines a structure diagram that shows the classes, operations, 

attributes, and associations of the model. Class diagrams highlight artifacts (classes) that 

create objects based on common (shared) operations, relationships, and attributes 

(Alhumaidan, 2012). The same is outlined in figure 4.6 below. 



 
 

40 

 

 

Figure 4. 6 Class Diagram 

 

4.3.6 Wireframes of the Mobile Application  

Figure 4.7 and 4.8 below shows the mobile application’s wireframes. The user interface 

is designed to be simple and easy to use. The user has to register and log in to upload the 

image and retrieve the venom category of the snake and its recommended antidote. 

 



 
 

41 

 

 

Figure 4. 7 Login Wireframes 



 
 

42 

 

 

Figure 4. 8 Classification Wireframes 



 
 

43 

 

Chapter 5: Model Development, Implementation, and Testing 

5.1 Introduction 

This chapter described the layout of the system. The model implementation stage 

described data preparation, the various steps that were taken to build the model, and its 

testing. In addition, the programming tools, and development environment are all 

discussed in this chapter. As described in the conceptual model in the literature reviewed 

in chapter two, the research entailed retraining, validating, and testing the MobileNetV2 

model using Kenyan Snake Image datasets. The resulting model was embedded into the 

Android mobile application. Testing included functional and usability tests to check if the 

developed model accomplished the objectives set out at the beginning of the research 

project. This study aimed to build a snake species classification model for snakebite 

envenomation management. 

5.2 Development environment and Language 

The development of the model was carried out on the Google Colab platform. Google 

Colab is a free cloud service that offers free GPU (Graphics Processing Unit) that were 

essential for building this model. A GPU performs matrix operations like convolution 

much more quickly than a CPU (LeCun, Bengio, & Hinton, 2015). The application of 

deep learning required very expensive hardware and powerful libraries such as Keras, 

OpenCV, sklearn, and TensorFlow. The Google Colab platform was more preferred since 

it provides access to powerful libraries such as OpenCV, Keras, and TensorFlow which 

were essential in developing the model. It was also possible to automatically back up the 

notebooks containing the code and the image datasets on the google drive. All training 

and validations were used in the same environment. 

5.2.1 Software Requirements 

The software resources consisted of the cloud-based hardware components used on the 

Google Colab development environment. There was limited control over these resources 

as they were being accessed on a Google Colab which is a cloud-based platform. Python 

language was preferred because of its ease of use and compatibility with the Colab. The 

table below highlighted the software requirements of this research. 



 
 

44 

 

Table 5. 1 Software Requirements 

Software Details 

Google Cloud GPU GPU 1xTesla K80, having 2496 CUDA cores, compute 

3.7, 12GB (11.439GB Usable) GDDR5 VRAM 

Python 3.10 Libraries Version 

 Tensorflow 2.8 

 Tonsorboard 2.8 

 Numpy 1.21.2 

 Keras 1.16.2 

 Augmentor 8.0.1 

 PIL 8.0.1 

 Scikit Learn 0.24 

 Cv2 4.5.1 

 Matplotlib 3.3.4 

 Split-Folder 0.4.3 

Kotlin/Java XML & 

Groovy 

 

Adobe Lightroom  3.7.0 

Android Studio  4.5 

SQL lite  2.2.0 

5.2.2. Hardware Requirements 

The system was developed in the following hardware environment. 

i. HP Pavilion 15Notebook PC, 64-BIT operating system, x64-based processor 

ii. Intel(R) Core (TM) i7-5500U CPU @ 2.30GHz, 2.40 GHz  

iii. 12.0 GB installed RAM 

iv. 500GB Hard Disk  

v. Microsoft Windows 10 pro  



 
 

45 

 

5.3 Model Components 

The model was created using transfer learning. Features extracted from the activation of 

the MobileNetV2 deep neural network, which has been trained on a large, fixed set of 

image classification tasks, were re-purposed specifically for this project. As reviewed in 

Chapter two, MobileNetV2 was chosen instead of bigger networks like DensNet because 

the MobileNetV2 is specifically designed for on-device vision tasks like image 

classification. They create smaller, fast models yet do not lose much in terms of accuracy 

(Sandler & Howard, MobileNetV2: The Next Generation of On-Device Computer Vision 

Networks, 2018). The MobileNet