Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

370 
https://internationalpubls.com 

Integration of Mathematical Operators Based on Decision 

Boundary Complexity and Combinatorial Optimization for 

Improved Deep Learning Classifiers 
 

Shruti Thapar1, Krati Sharma2, Dharmveer Yadav3
, Budesh Kanwer4

, Ashish Raj5 

1Department of Electronics & Communication, Poornima Institute of Engineering & Technology, Jaipur, 

Rajasthan, India 

2Department of English and Soft Skills, Poornima Institute of Engineering & Technology, Jaipur, Rajasthan, 

India 
3Department of Computer Science, St. Xavier’s College, Jaipur, Rajasthan, India 

4Department of AI&DS, Poornima Institute of Engineering & Technology, Jaipur, Rajasthan, India 

5Department of Electrical and Electronics Engineering, Poornima University, Jaipur, Rajasthan, India 

Email- shruti.thapar@poornima.org 1, kratibhomia@gmail.com 2, dharmveerya21@gmail.com 3, 

budesh82@gmail.com 4, ashishraj1987@gmail.com 5 

 

Article History: 

Received: 16-02-2024 

Revised: 22-04-2024 

Accepted: 05-05-2024 

Abstract:  

 The proliferation of complex diseases in livestock, such as lumpy skin disease, 

demands advanced diagnostic tools that can accurately classify and predict outbreaks. 

This study explores the integration of complex mathematical operators within deep 

learning classifiers to enhance their accuracy and efficiency in diagnosing lumpy skin 

disease. By focusing on the decision boundary complexity, which delineates different 

disease states in high-dimensional spaces, and employing combinatorial optimization 

techniques, we develop a novel framework that significantly improves classification 

performance. The methodology hinges on optimizing the configuration and 

combination of mathematical operators, such as gradient operators and higher-order 

derivatives, to refine feature extraction processes. This approach allows for a more 

nuanced understanding of the disease features that are critical for accurate classification. 

Using a dataset comprised of clinical and image data from infected cattle, our enhanced 

classifiers demonstrate a marked improvement in predictive accuracy compared to 

traditional deep learning models. The case study not only underscores the potential of 

integrating advanced mathematical concepts into deep learning but also sets a precedent 

for tackling similar challenges in veterinary medicine and beyond. 

Keywords: Deep Learning Classifiers, Lumpy Skin Disease, Decision Boundary 

Complexity, Combinatorial Optimization, Mathematical Operators, Feature Extraction, 

Livestock Disease Classification, Predictive Accuracy, Veterinary Medicine, High-

Dimensional Data Analysis. 

 

1. MATHEMATICAL FOUNDATIONS AND OPTIMIZATION TECHNIQUES:  

Deep learning, a subset of machine learning, has revolutionized numerous fields from image 

and speech recognition to autonomous driving and medical diagnosis. At its core, deep learning 

constructs algorithms, known as neural networks, that can learn and make intelligent decisions 

on their own. This ability stems from the network’s capacity to perform complex mathematical 

operations and optimize these operations for better performance and accuracy. Understanding 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

371 
https://internationalpubls.com 

these underlying mathematical concepts and optimization techniques is crucial for advancing 

deep learning applications. 

Mathematical Operators in Neural Networks 

The foundation of any neural network is its ability to perform mathematical calculations on 

input data to extract features and make predictions. These calculations are primarily linear 

algebraic operations, which involve vectors, matrices, and tensors. At the basic level, a neural 

network transforms its input through a series of layers, each defined by a set of weights and 

biases, and applies nonlinear transformations to enable complex representations. 

1. Linear Transformations and Decision Boundaries: Every neural network layer 

essentially performs a linear transformation followed by a nonlinear activation. The 

simplest form of a linear transformation in a binary classifier can be represented by the 

decision boundary equation 𝑓(𝑥)=𝑤𝑇𝑥+𝑏=0 .This equation defines a hyperplane that 

divides the input space into two decision areas, each corresponding to one of the classes. 

2. Optimization Techniques: Optimization is at the heart of training deep learning 

models. It involves adjusting the weights of the network to minimize a loss function, 

which measures the difference between the predicted output and the actual output. The 

Gradient Descent Update Rule 𝑤𝑛𝑒𝑤=𝑤𝑜𝑙𝑑−𝛼∇𝑤𝐿wnew=wold−α∇wL is a primary 

example of how weights are iteratively adjusted in the direction that most reduces the 

loss. 

Activation Functions 

Activation functions introduce non-linear properties to the network which allows the model to 

learn more complex patterns. Two common examples are: 

3. Sigmoid Activation Function: 𝜎(𝑧)=11+𝑒−𝑧 is particularly useful in binary 

classification as it maps any real-valued number into the (0,1)(0,1) interval, 

representing a probability-like output. 

4. Softmax Function for Multi-Class Classification: For scenarios where classifications 

are mutually exclusive, the SoftMax function normalizes the outputs, converting them 

into probability scores summing to one. 

Loss Functions 

To quantify how well a network performs, loss functions are crucial. They provide a metric to 

evaluate the errors made by a network during training. 

5. Cross-Entropy Loss: In classification problems, cross-entropy helps measure the 

performance by calculating the total entropy between the predicted probabilities and 

the actual distribution. 

Regularization Techniques 

Overfitting is a common problem where a model learns the detail and noise in the training data 

to an extent that it negatively impacts the performance of the model on new data. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

372 
https://internationalpubls.com 

6. L2 Regularization: Also known as Ridge Regression, it adds a penalty on the size of 

coefficients. 𝑅(𝑤)=𝜆∥𝑤∥2R(w)=λ∥w∥2 where 𝜆λ is a complexity parameter that 

controls the amount of shrinkage: the larger the value of 𝜆λ, the greater the amount of 

shrinkage. 

7. Dropout: A different type of regularization where randomly selected neurons are 

ignored during training, reducing the chance for overfitting. 

Advanced Optimization 

Beyond simple gradient descent, more sophisticated algorithms can improve convergence. 

8. Stochastic Gradient Descent (SGD) and Adam Optimizer: While SGD updates 

parameters with a subset of training data rather than the full dataset for faster 

computations, Adam, on the other hand, computes adaptive learning rates for each 

parameter. 

Beyond Classical Layers: Specialized Operations 

Deep learning also involves more complex operations which are suited for specific data types 

and problems. 

9. Backpropagation Through Time in RNNs: Essential for training Recurrent Neural 

Networks, it involves moving backward through the network to calculate gradients over 

time, crucial for sequence prediction problems. 

10. Convolution Operations in CNNs: Used predominantly in processing image data, this 

operation extracts features from local image patches through a filter. 

Decision Boundary Equation: 

𝑓(𝑥) = 𝑤𝑇𝑥 + 𝑏 − 0                                                                                      (1) 

This equation defines the decision boundary of a linear classifier, where 𝑤 is the weight vector, 

𝑥 is the input feature vector, and 𝑏 is the bias. The decision boundary separates different classes 

in the feature space. 

Gradient Descent Update Rule: 

𝑤new = 𝑤old − 𝛼∇𝑤𝐿                                                                                    (2) 

This equation represents the update rule for the weights 𝑤 in gradient descent optimization, 

where 𝛼 is the learning rate and ∇𝑤𝐿 is the gradient of the loss function 𝐿 with respect to the 

weights. 

Sigmoid Activation Function: 

𝜎(𝑧) =
1

1+𝑒−𝑧
                                                                                                  (3) 

The sigmoid function is used as an activation function in neural networks to map real valued 

inputs to the (0,1) range, facilitating binary classification tasks. 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

373 
https://internationalpubls.com 

SoftMax Function for Multi-class Classification: 

Softmax⁡(𝑧𝑖) =
𝑒2

2

∑  𝑖  𝑒
3
                                                                                    (4) 

SoftMax function is used in multi-class classification to convert the output logits 𝑧 into 

probabilities by normalizing them across all classes. 

Cross-Entropy Loss for Binary Classification: 

𝐿(𝑦, �̂�) = −[𝑦log⁡(�̂�) + (1 − 𝑦)log⁡(1 − �̂�)]                                            (5) 

Cross-entropy loss measures the performance of a classification model whose output is a 

probability value between 0 and 1 . It penalizes the deviation from actual labels 𝑦 and predicted 

probabilities �̂�. 

L2 Regularization Term: 

𝑅(𝑤) = 𝜆 ∥ 𝑤 ∥2                                                                                       (6) 

L2 regularization is used to prevent overfitting by adding a penalty term 𝑅(𝑤) to the loss 

function, which is proportional to the square of the magnitude of the weights, controlled by the 

regularization parameter 𝜆. 

Stochastic Gradient Descent (SGD): 

𝑤 = 𝑤 − 𝛼∇𝑤𝐿𝑖                                                                                        (7) 

Stochastic Gradient Descent updates the model weights using the gradient of the loss function 

𝐿𝑖 with respect to a randomly selected subset of data, improving optimization speed. 

 

Backpropagation Through Time (For RNNs): 

∂𝐿

∂𝑤
= ∑  𝑡

∂𝐿𝑢

∂𝑤
                                                                                              (8) 

This equation is used to calculate the gradients of loss function 𝐿 with respect to weights 𝑤 in 

recurrent neural networks ( RNNs ), summed over all time steps 𝑡. 

Convolution Operation in CNNs: 

(𝑓 ∗ 𝑔)(𝑡) = ∫ ⁡ 𝑓(𝜏)𝑔(𝑡 − 𝜏)𝑑𝜏                                                                 (9) 

In the context of convolutional neural networks (CNNs), this equation describes the 

convolution operation between an input function 𝑓 and a kernel 𝑔, used for feature extraction. 

Max Pooling in CNNs: 

MaxPooling⁡(𝐴) = max
𝑖,𝑗∈𝐴

 𝑎𝑖𝑗                                                                         (10) 

Max pooling reduces the dimensionality of each feature map in CNNs by taking the maximum 

value over a specified window 𝐴. 

This selection includes equations relevant to various aspects of deep learning, from basic 

operations to more complex functions involving optimization and regularization.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

374 
https://internationalpubls.com 

Euler's Formula for Complex Exponentials: 

𝑒𝑖𝜃 = cos⁡(𝜃) + 𝑖sin⁡(𝜃)                                                                                (11) 

This fundamental equation in complex analysis is used for transforming complex numbers into 

their exponential form, which can be useful in signal processing aspects of neural networks. 

Laplacian Operator in Graph Theory: 

𝐿 = 𝐷 − 𝐴                                                                                                      (12) 

The Laplacian matrix 𝐿 of a graph is used in graph neural networks, where 𝐷 is the degree 

matrix and 𝐴 is the adjacency matrix, helping to define the structure of the graph. 

Eigenvalue Decomposition for PCA: 

𝑋𝑇𝑋𝑣 = 𝜆𝑣                                                                                                  (13) 

Principal Component Analysis (PCA) involves finding the eigenvalues 𝜆 and eigenvectors 𝑣 

of the covariance matrix 𝑋𝑇𝑋 to reduce dimensions and extract features. 

ReLU Activation Function: 

ReLU⁡(𝑥) − max(0, 𝑥) 

The Rectified Linear Unit (ReLU) activation function is widely used in deep learning for 

introducing non-linearity, improving training speeds and convergence. 

Adam Optimizer Update Rule: 

𝑤𝑡+1 = 𝑤𝑡 −
𝛼

√6𝑡+𝑒
�̂�𝑡                                                                                (14) 

Adam optimization algorithm uses adaptive learning rates for each parameter by estimating 

first ( 𝑚 ) and second (v) moments of the gradients, improving the efficiency of network 

training. 

Batch Normalization Transform: 

𝑦 = 𝛾 (
𝑥−𝜇

√𝜎2+𝜀
) + 𝛽                                                                                    (15) 

Batch normalization stabilizes and accelerates deep network training by normalizing the inputs 

(layer activations) 𝑥 using mean 𝜇 and variance 𝜎2, scaled by 𝛾 and shifted by 𝛽. 

Hinge Loss for SVM: 

𝐿 = max(0,1 − 𝑦 ⋅ 𝑓(𝑥))                                                                                       (16) 

Used primarily in support vector machines (SVMs) for classification, hinge loss penalizes 

misclassifications and instances where the classification margin is not maintained. 

Entropy in Information Theory: 

𝐻(𝑋) = −∑  𝑖 𝑝(𝑥𝑖)log⁡ 𝑝(𝑥𝑖)                                                                                (17) 

Entropy measures the uncertainty or impurity in a dataset and is used in various machine 

learning algorithms to gauge information gain and decision making. 

Gaussian Kernel in SVMs: 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

375 
https://internationalpubls.com 

𝐾(𝑥, 𝑥′) = exp⁡ (−
∥∥𝑥−𝑥′∥∥

2

2𝜎2
)                                                                                  (18) 

The Gaussian or RBF kernel is used in SVMs for transforming the input space into a higher-

dimensional space, allowing for the handling of nonlinear boundaries. 

Dropout Regularization: 

𝑟𝑗 ∼ Bernoulli⁡(𝑝)

𝑥‾ = 𝑥 ∗ 𝑟
                                                                                        (19) 

Dropout prevents overfitting in neural networks by randomly setting a fraction 𝑝 of input units 

to zero at each update during training. 

Attention Mechanism in Transformers: 

Attention (𝑄, 𝐾, 𝑉) = softmax⁡ (
𝑄𝐾𝑇

𝑉𝑑𝑘
) 𝑉                                                             (20) 

In the context of transformer models, this equation calculates the attention weights and outputs 

by scaling the dot products of queries 𝑄 and keys 𝐾 with values 𝑉. 

Gated Recurrent Unit (GRU) Update Gate: 

𝑧𝑡 = 𝜎(𝑊𝑧 ⋅ [ℎ𝑡−1, 𝑥𝑡])                                                                                       (21) 

The update gate in GRU helps the model to determine how much of the past information 

(from previous time steps) needs to be passed along to the future. 

Gated Recurrent Unit (GRU) Reset Gate: 

𝑟𝑡 = 𝜎(𝑊𝑟 ⋅ [ℎ𝑡−1, 𝑥𝑡])                                                                                      (22) 

The reset gate in GRU controls how much of the past information to forget, which helps in 

capturing dependencies and managing vanishing gradient issues in sequences. 

Gated Recurrent Unit (GRU) Current Memory Content: 

ℎ‾𝑡 = tanh⁡(𝑊 ⋅ [𝑟𝑡 ∗ ℎ𝑡−1, 𝑥𝑡])                                                                             (23)    

This represents the candidate activation or new memory content in GRU, combining the reset 

gate's output with the current input. 

Gated Recurrent Unit (GRU) Final Memory at Current Step: 

ℎ𝑡 = (1 − 𝑧𝑡) ∗ ℎ𝑡−1 + 𝑧𝑡 ∗ ℎ‾𝑡                                                                                (25) 

This final state for GRU at time 𝑡 blends the old state ℎ𝑡−1 and the new candidate state ℎ‾𝑡 based 

on the update gate 𝑧𝑡. 

L-BFGS Update Formula for Optimization: 

𝑠𝑘 = 𝑥𝑘+1 − 𝑥𝑘⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(26)

𝑦𝑘 = ∇𝑓(𝑥𝑘+1) − ∇𝑓(𝑥𝑘)⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(27)

𝜌𝑘 =
1

𝑦
𝑘
𝑇𝑠𝑘

⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(28)

𝐻𝑘+1 = (𝐼 − 𝜌𝑘𝑠𝑘𝑦𝑘
𝑇)𝐻𝑘(𝐼 − 𝜌𝑘𝑦𝑘𝑠𝑘

𝑇) + 𝜌𝑘𝑠𝑘𝑠𝑘
𝑇⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(29)

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

376 
https://internationalpubls.com 

L-BFGS is a quasi-Newton method that approximates the Hessian matrix involved in the 

optimization of nonlinear functions, suitable for large-scale problems in deep learning. 

Tikhonov Regularization (Ridge Regression): 

min
𝑤

  ∥ 𝑋𝑤 − 𝑦 ∥2+ 𝜆 ∥ 𝑤 ∥2                                                                                   (30) 

Tikhonov regularization, or ridge regression, is used to solve regression problems by imposing 

a penalty on the size of coefficients. 

Elastic Net Regularization: 

min
𝑤

  ∥ 𝑋𝑤 − 𝑦 ∥2+ 𝜆1 ∥ 𝑤 ∥1+ 𝜆2 ∥ 𝑤 ∥2                                                               (31) 

Elastic net combines L1 and L2 regularization, helping in variable selection (L1) and shrinking 

coefficients (L2), beneficial in scenarios with highly correlated variables. 

Kullback-Leibler Divergence for Probability Distributions: 

𝐷𝐾𝐿(𝑃 ∥ 𝑄) = ∑  𝑖 𝑃(𝑖)log⁡
𝑃(𝑖)

𝑄(𝑖)
                                                                                  (32) 

Kullback-Leibler divergence measures how one probability distribution diverges from a 

second, expected probability distribution. 

Dynamic Programming Recurrence Relation: 

𝑉(𝑖, 𝑤) = max(𝑉(𝑖 − 1,𝑤), 𝑉(𝑖 − 1,𝑤 − 𝑤𝑖) + 𝑣𝑖)                                              (33) 

This equation is used in the knapsack problem, a common problem in combinatorial 

optimization, where 𝑉(𝑖, 𝑤) is the maximum value that can be achieved with the first 𝑖 items 

and a weight limit 𝑤. 

These equations collectively illustrate the mathematical and computational complexity 

involved in designing and optimizing deep learning models, particularly when dealing with 

sophisticated classifiers and large datasets. The equations and concepts discussed represent the 

backbone of deep learning technologies. By integrating these mathematical principles, deep 

learning models can be optimized and tailored to specific problems more effectively. As we 

continue to push the boundaries of what is possible with these models, understanding and 

applying these foundational equations remain critical.Each component, from linear 

transformations and regularization to loss functions and specialized neural network operations, 

plays a crucial role in building robust, efficient, and scalable models. The continuous evolution 

of these mathematical frameworks is what will drive future innovations in deep learning, 

enabling smarter and more intuitive artificial intelligence systems. Whether for simple 

applications or complex, domain-specific tasks, the integration of complex mathematical 

operators and advanced optimization techniques will remain at the forefront of technological 

advancements in AI. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

377 
https://internationalpubls.com 

 

Figure 1. Different Samples of Foot-Mouth Disease and Skin Disease  

The introduction of these technologies underscores a larger trend toward the automation and 

enhancement of disease detection methods in agriculture. Deep learning, through the use of 

convolutional neural networks (CNNs), enables the analysis of cattle images to identify and 

classify diseases accurately. This not only aids in early detection but also enhances the overall 

management of cattle health, contributing to more sustainable agricultural practices. Our 

research aims to further these advancements by focusing on the development of an efficient 

cattle disease detection system that is accessible and user-friendly. By leveraging deep learning 

and smartphone technology, our proposed system aims to empower farmers and veterinarians 

to diagnose diseases accurately and efficiently, even in resource-constrained environments. 

This approach addresses the critical problem of delayed or inaccurate disease diagnosis, which 

can lead to prolonged animal suffering, heightened economic losses, and increased risk of 

disease spread. In essence, the problem our research seeks to solve is the significant gap in 

timely and accurate disease detection in cattle, which can be transformative for cattle farming. 

By employing advanced AI technologies and making these tools accessible through 

smartphones, we can revolutionize how diseases are managed in the livestock industry. The 

following sections of this paper will detail the methodology, implementation, evaluation 

results, and potential impact of our system, aiming to provide a comprehensive solution to a 

pressing issue in global agriculture. 

2. RELATED STUDY 

Cattle farming, a cornerstone of global agriculture, faces significant challenges due to diseases 

that can severely impact livestock health and productivity. This literature review explores the 

recent advancements in machine learning and deep learning techniques applied to cattle disease 

detection, classification, and management. It draws on various studies to highlight the current 

technologies, methodologies, and the effectiveness of these approaches, as well as the existing 

gaps and future directions in research.  

 

 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

378 
https://internationalpubls.com 

Table 1. Analysis of Methodologies 

Reference Research 

Purpose 

Methodologies Used Key Findings Challenges/Research 

Gaps 

Y. Li et al.  Classification of 

cattle behaviors 

KNN, Random Forest, 

Extreme Boosting 

High accuracy in 

behavior 

classification 

Limited temporal 

information affects 

accuracy 

K. Liakos et 

al.  

Real-time 

lameness 

detection 

ANN, SVM, Random 

Forest 

Effective in 

lameness 

detection 

Needs more reliable and 

practical systems 

D. Wu et al.  Detection of cattle 

lameness 

YOLOv3, SVM, 

KNN, DTC 

High accuracy in 

lameness 

detection 

Limited datasets 

M. E. Pastell 

et al.  

Lameness 

detection using 

PNN 

Probabilistic Neural 

Network (PNN), 4-

balance system data 

High accuracy in 

identifying 

lameness 

System may not be 

suitable for all farms 

N. V. 

Kishan et al.  

Health monitoring 

and disease 

prediction 

Data mining, k-NN, 

Naive Bayes, SVM, 

IoT 

Focus on rare 

diseases and 

veterinary care 

Integration with existing 

practices 

G. Rai et al.  Prediction of 

Lumpy Skin 

Disease 

CNN, ANN, VGG-16, 

VGG-19, Inception-v3 

High accuracy in 

disease prediction 

No standard dataset 

available 

Y. Qiao et 

al. 

Multi-cattle 

segmentation 

Mask R-CNN High accuracy in 

segmentation 

Superior to traditional 

segmentation methods 

R. Dulal et 

al.  

Cattle 

identification 

using YOLOv5 

YOLOv5 Effective in 

overcoming RFID 

limitations 

Cost and scalability 

analysis needed 

 

Y. Li et al. developed a method for classifying multiple cattle behaviors using machine learning 

algorithms like K-nearest neighbors and random forests, achieving a high F1 score accuracy. 

However, their study noted limitations in capturing temporal information which affected the 

accuracy of movement classifications such as feed tossing and rolling biting . Several 

researchers have focused on detecting lameness, a prevalent issue in cattle which can 

significantly affect their welfare and farm productivity.  K. Liakos et al. trained Artificial 

Neural Networks (ANN) and used Support Vector Machines (SVM) for real-time lameness 

detection. This approach highlighted the potential of machine learning to improve the accuracy 

and reliability of automated lameness detection systems . D. Wu et al. also targeted lameness 

detection using a combination of YOLOv3, SVM, K-Nearest Neighbor, and Decision Tree 

Classifiers. Their approach achieved a notable accuracy of 98.57%, demonstrating the efficacy 

of deep learning algorithms in diagnosing complex conditions like lameness .M. E. Pastell et 

al. utilized a Probabilistic Neural Network (PNN) model using data from a 4-balance system 

to detect lameness with high accuracy. This system, while effective, was noted to be potentially 

unsuitable for all farm types due to its design and implementation challenges .N. V. Kishan et 

al. explored the use of data mining techniques alongside IoT-based systems for cattle health 

monitoring and disease prediction. Their approach used classification methods such as k-

nearest neighbors, Naive Bayes, and SVM to focus on rare diseases and improve veterinary 

practices. G. Rai et al. applied convolutional neural networks (CNN) like VGG-16 and 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

379 
https://internationalpubls.com 

Inception-v3 for predicting Lumpy Skin Disease, achieving a high accuracy of 92.5%. 

However, the absence of a standard dataset posed a significant challenge . Y. Qiao et al. 

introduced a method for multi-cattle segmentation using Mask R-CNN, achieving a mean pixel 

accuracy of 92%. This method surpassed traditional instance segmentation methods like 

SharpMask and DeepMask, highlighting the potential of deep learning in precise livestock 

management . R. Dulal et al. used the YOLOv5 model for cattle identification, focusing on 

overcoming the limitations of RFID systems through deep learning techniques. This study 

provided insights into cost-effective and scalable solutions for cattle management. The 

integration of advanced technologies in real-world farming operations remains a significant 

challenge. The need for systems that are not only accurate but also user-friendly and adaptable 

to different farming conditions is critical for widespread adoption [10-19]. 

3. MATERIALS AND METHODS 

This research focuses on creating an advanced system for detecting cattle diseases using deep 

learning techniques, particularly Convolutional Neural Networks (CNNs) and transfer 

learning. The primary goal is to assist farmers and veterinarians in the early detection and 

diagnosis of cattle diseases, thereby enhancing animal health and productivity. The 

methodology utilizes state-of-the-art deep learning models, known for their effectiveness in 

image classification, tailored specifically for identifying various cattle diseases from visual 

cues[20]. 

 

Figure 2. Process Flow Diagram 

Figure 2 displays the workflow from data acquisition to the output of results, outlining the 

steps involved in processing and analyzing the images. Figure 3 provide insights into the 

distribution of images across different classes, highlighting the diversity and balance of the 

dataset which is crucial for training unbiased models.Figure 4 shows a pie chart that represents 

the proportion of images per disease category, emphasizing the dataset's variety and the need 

for a model capable of recognizing a broad spectrum of disease indicators. 

Dataset Visualization and Acquisition 

The foundation of any deep learning application lies in the quality and structure of the dataset 

used. For this study, a comprehensive dataset consisting of 455 images was compiled. These 

images include various cattle diseases and healthy specimens, categorized into different classes 

based on the disease type. Each class is distinctly labeled to facilitate supervised learning, 

where the model learns to associate specific images with their corresponding disease labels.  

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

380 
https://internationalpubls.com 

Data Pre-processing 

Data pre-processing is critical to enhance the data quality, which in turn, improves the learning 

efficiency and performance of the model. 

Data Noise Removal: Noise reduction is pivotal in image processing, especially for medical 

and biological imaging where clarity and detail are essential for accurate diagnosis. Techniques 

were implemented to reduce noise and remove irrelevant background elements, enhancing the 

contrast and focus on key features essential for disease classification[22]. 

 

Figure 3. Sample Distribution of Database 

Data Augmentation: To bolster the model's robustness and its ability to generalize across 

different visual representations of cattle diseases, various data augmentation techniques were 

applied. These included geometric transformations like rotations, flips, and translations; color 

adjustments to mimic different lighting conditions; and zooms and crops to vary the image 

focus and perspective. Additionally, artificial noise and blur were introduced to ensure the 

model's effectiveness under varied photographic conditions, simulating real-world scenarios 

where images may not always be perfect. 

Deep Learning Models 

The project utilizes several advanced deep learning models, adapting their architecture through 

the process of transfer learning to meet the specific requirements of cattle disease detection. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

381 
https://internationalpubls.com 

 

Figure 4. Design of VGG 16 Based Classifier 

VGG16: VGG16, known for its depth and simplicity, was selected for its proven track record 

in image classification tasks. The model, pre-trained on the ImageNet database, was fine-tuned 

on the cattle disease dataset. The last few layers were specifically modified to align with the 

number of disease classes identified in the dataset. This fine-tuning allows the pre-trained 

network to adapt its learned features to new, specific tasks of recognizing cattle diseases[23-

25]. 

 

Figure 5. Design of Inception V3 Based Classifier 

Inception V3: Inception V3 is another model chosen for its sophisticated architecture designed 

to capture information at multiple scales using various convolutions and pooling strategies. The 

model was fine-tuned with additional layers like batch normalization and dropout to enhance 

training stability and minimize overfitting, adapting it to the nuances of cattle disease imaging. 

ResNet50: ResNet50 employs residual learning to facilitate the training of very deep networks 

by incorporating skip connections that jump over some layers. These connections help mitigate 

the vanishing gradient problem, allowing for deeper network architectures without a loss in 

performance. ResNet50 was tailored for deeper and more precise feature extraction, offering 

substantial computational efficiency over other deep networks without residual connections 

[27]. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

382 
https://internationalpubls.com 

 

Figure 6. Distribution of Diseases in Database 

Training Details: 

• Optimization Algorithms: The Adam optimizer was used due to its adaptive learning 

rate features, enhancing the convergence speed. 

• Loss Functions: Cross-entropy loss, a standard for multiclass classification tasks, was 

employed to compute the model's loss. 

• Performance Metrics: Metrics such as Accuracy, Precision, Recall, and F1-Score 

were meticulously recorded to evaluate the model's performance and ensure reliability. 

 

Figure 7. Analysis of Distribution of Diseases in Database 

Evaluation: 

• Validation Strategy: The dataset was strategically divided into training, validation, 

and test sets, ensuring that the model's performance is rigorously evaluated on unseen 

data. 

• Cross-validation: Used to ascertain the consistency and reliability of the model across 

different data subsets, ensuring that the model performs well in various scenarios and 

setups. 

This research methodology leverages the powerful capabilities of CNNs and transfer learning 

to develop a robust, accurate, and efficient system for cattle disease detection. By combining 

sophisticated image processing techniques with cutting-edge deep learning technologies, this 

system is poised to make significant contributions to veterinary medicine. It promises to 

enhance the capabilities for early disease detection and management in cattle, thereby 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

383 
https://internationalpubls.com 

improving animal health, increasing farm productivity, and reducing economic losses 

associated with cattle diseases [28-32]. 

4. RESULTS & DISCUSSION 

Recent advances in AI, particularly the development of deep learning technologies, have 

opened up new avenues for innovation in disease detection. CNNs, a class of deep neural 

networks, are particularly suited for image recognition tasks and have been successfully applied 

in various fields including medical imaging and autonomous driving. These networks 

automatically learn to detect complex patterns in data, making them ideal for recognizing signs 

of disease in images of cattle. The primary objective of this study was to develop a robust 

automated system capable of detecting specific cattle diseases from images using CNN 

architectures such as InceptionV3, MobileNetV2, and DenseNet169. This system aimed to 

classify images into distinct categories based on the presence of foot, mouth, and lumpy skin 

diseases, leveraging a dataset compiled from various online sources. A comprehensive dataset 

was essential for training and testing the deep learning models. The dataset consisted of images 

classified into different disease categories, providing a varied pool of data points for the models 

to learn from. The images were collected from openly available internet sources, ensuring a 

diverse representation of disease manifestations. This diversity is critical in training robust 

models that can generalize well across different real-world scenarios. The study utilized the 

Python programming language, known for its extensive libraries and frameworks that facilitate 

data handling, processing, and model development. Key technologies used included 

TensorFlow and Keras for building and training the CNN models, and NumPy for performing 

mathematical operations essential for data preprocessing and analysis. 

The performance of each model was evaluated using standard metrics derived from the 

confusion matrix, including accuracy, precision, recall, and the F1-score. These metrics 

provide a comprehensive understanding of model performance, highlighting strengths and 

weaknesses in specific areas such as the model’s ability to correctly identify disease-positive 

cases (precision) and its sensitivity in detecting these cases among all positive instances 

(recall). 

Classification Accuracy: The accuracy of classification is determined by determining the 

percentage of data points that have been successfully classified out of the total number of data 

points. Calculated by taking the total number of data points and dividing it by the sum of true 

positives (TP) and true negatives (TN), it includes the following:  

Accuracy=
𝐓𝐏+𝐓𝐍+𝐅𝐏+𝐅𝐍

𝐓𝐏+𝐓𝐍
                                                                                            (34) 

Precision: Accuracy can be defined as the ratio of the actual positive to the total number of 

positives anticipated. A model with a high precision has a low false positive rate, which means 

that it only sometimes incorrectly identifies negative occurrences as positive. Conversely, a 

low precision suggests that the model tends to make a significant number of false positive 

predictions, which can be problematic in scenarios where false positives are costly or 

undesirable as in fraud detection, and spam filtering. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

384 
https://internationalpubls.com 

                               Precision =  
𝐓𝐏

𝐓𝐏+𝐅𝐏
                                             (35) 

Recall, - In classification tasks, a performance indicator that is also known as sensitivity or 

true positive rate is utilized to evaluate the capability of a model to accurately identify all 

positive cases from the entire number of actual positive examples that are contained inside the 

dataset. A formula that is used to compute it is as follows: 

Recall-=
𝐓𝐏

𝐓𝐏+𝐅𝐍
                                                                                                              (36) 

 

 
Figure 8. Performance Parameters of Classifier 

 
Figure 9. Comparative Analysis of Performance Parameters of Classifier 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

385 
https://internationalpubls.com 

 

Figure 10. Comparative Analysis of Confusion Matrix for VGG-16 and Combined Classifier 

In the contemporary landscape of agricultural technology, the implementation of artificial 

intelligence (AI), particularly deep learning models, offers a promising frontier for enhancing 

livestock disease detection. This analysis focuses on evaluating the performance of various 

convolutional neural networks (CNNs) employed in the detection of cattle diseases, including 

Inception V3, MobileNetV2, DenseNet169, VGG-16, and RESNET-50. The study uses a 

tailored dataset that reflects common cattle ailments, such as foot, mouth, and lumpy skin 

diseases. The dataset comprises images sourced from various online repositories, classified into 

specific disease categories to train the CNN models. This approach is designed to simulate a 

real-world application where visual data from livestock would be analyzed to predict disease 

presence. The models were evaluated based on several metrics, including training and 

validation accuracy, precision, recall, and F1-score, which provide a comprehensive view of 

model performance.The results presented in the study are as follows: 

• Inception V3: Demonstrated high reliability with a training accuracy of 97.60% and 

validation accuracy of 97.98%. The model showed robust precision and recall, making 

it suitable for scenarios where both false positives and false negatives carry significant 

consequences. 

• MobileNetV2: Featured slightly lower validation accuracy than Inception V3, 

indicating potential overfitting or sensitivity to the dataset's variability. Nevertheless, 

its high precision and recall suggest it remains an effective tool for disease detection. 

• DenseNet169: Emerged as the top performer with the highest validation accuracy 

(99.37%) and excellent precision and recall. Its architecture possibly offers better 

feature extraction capabilities, which is crucial for detailed image analysis in disease 

identification. 

• VGG-16: While it had the lowest validation accuracy among the top performers, its 

precision and recall were high. This model might be more sensitive to the training 

dataset's nuances and could benefit from further tuning. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

386 
https://internationalpubls.com 

• RESNET-50: Although it had the highest training accuracy (99.40%), its lower 

validation accuracy and significantly lower precision and recall indicate issues with 

generalizability and potential overfitting. 

• Combined: This model, likely an ensemble of different architectures, showed balanced 

performance across all metrics, demonstrating the utility of model ensembling to 

stabilize detection performance. 

Comparing these results with historical data from other studies underscores the advancement 

in deep learning applications for agricultural uses. The DenseNet169 model, in particular, 

stands out by setting a high benchmark in validation accuracy and F1-score, indicating superior 

generalization over unseen data, a critical factor for practical deployment. 

The high performance of these CNN models suggests they could be integrated into mobile 

applications for real-time disease detection, providing farmers with timely, accurate 

diagnostics, thus allowing for quicker response and treatment. Future research could explore 

incorporating additional disease categories, enhancing data diversity, and deploying these 

models in field trials to further validate their effectiveness in practical setting. 

5. CONCLUSION 

In conclusion, the application of CNNs in detecting cattle diseases has demonstrated significant 

potential to revolutionize disease management in livestock. By continuing to refine these 

models and expand their capabilities, the agricultural sector can look forward to greater 

efficiency and accuracy in livestock healthcare, ultimately leading to improved productivity 

and reduced losses due to diseases. This study not only contributes to the field of agricultural 

AI but also sets the stage for future innovations that could further enhance the capabilities of 

AI in managing livestock health. The potential of AI to impact livestock disease management 

positively is immense and largely untapped. As demonstrated by the success of CNNs in this 

study, there is a clear path forward that involves technological innovation, careful consideration 

of ethical and practical concerns, and close collaboration between diverse stakeholders. By 

continuing to pursue these avenues, we can look forward to a future where livestock 

management is more data-driven, efficient, and responsive to the challenges posed by diseases. 

This will not only improve the health and productivity of livestock but also the economic 

viability of farms around the globe, securing food supply chains and enhancing the 

sustainability of agricultural practice. 

REFERENCES 

[1] G. Rai, Naveen, A. Hussain, A. Kumar, A. Ansari, and N. Khanduja, “A deep learning approach to detect 

lumpy skin disease in cows,” in Computer Networks, Big Data and IoT: Proceedings of ICCBI 2020.  

Springer, 2021, pp. 369–377. 

[2] M. Pastell and M. Kujala, “A probabilistic neural network model for lameness detection,” Journal of dairy 

science, vol. 90, no. 5, pp. 2283–2292, 2007. 

[3] G. Li, G. E. Erickson, and Y. Xiong, “Individual beef cattle identification using muzzle images and deep 

learning techniques,” Animals, vol. 12, no. 11, p. 1453, 2022. 

[4] N. Kreˇsi´c, I. Sˇimi´c, T. Bedekovi´c, Zˇ. Acinger-Rogi´c, and I. Lojki´c, “Evaluation of serological tests 

for detection of antibodies against lumpy skin disease virus,” Journal of Clinical Microbiology, vol. 58, no. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

387 
https://internationalpubls.com 

9, pp. 10–1128, 2020. 

[5] B. Lake, F. Getahun, and F. T. Teshome, “Application of artificial intelligence algorithm in image processing 

for cattle disease diagnosis,” Journal of Intelligent Learning Systems and Applications, vol. 14, no. 04, pp. 

71–88, 2022. 

[6] B. Jiang, Q. Wu, X. Yin, D. Wu, H. Song, and D. He, “Flyolov3 deep learning for key parts of dairy cow 

body detection,” Computers and Electronics in Agriculture, vol. 166, p. 104982, 2019. 

[7] V. R. Allugunti, “A machine learning model for skin disease classification using convolution neural 

network,” International Journal of Computing, Programming and Database Management, vol. 3, no. 1, pp. 

141–147, 2022. 

[8] A. Poursaberi, C. Bahr, A. Pluk, A. Van Nuffel, and D. Berckmans, “Real-time automatic lameness detection 

based on back posture extraction in dairy cattle: Shape analysis of cow with image processing techniques,” 

Computers and electronics in agriculture, vol. 74, no. 1, pp. 110–119, 2010. 

[9] E. Afshari Safavi, “Assessing machine learning techniques in forecasting lumpy skin disease occurrence 

based on meteorological and geospatial features,” Tropical Animal Health and Production, vol. 54, no. 1, p. 

55, 2022. 

[10] E. Arazo, R. Aly, and K. McGuinness, “Segmentation enhanced lameness detection in dairy cows from rgb 

and depth video,” arXiv preprint arXiv:2206.04449, 2022. 

[11] S. Shahinfar, D. Page, J. Guenther, V. Cabrera, P. Fricke, and K. Weigel, “Predic- tion of insemination 

outcomes in holstein dairy cattle using alternative machine learning algorithms,” Journal of dairy science, 

vol. 97, no. 2, pp. 731–742, 2014. 

[12] R. Chandrarathna, T. Weerasinghe, N. Madhuranga, T. Thennakoon, A. Gamage, and E. Lakmali, “’the 

taurus’: Cattle breeds & diseases identification mobile application using machine learning,” arXiv preprint 

arXiv:2302.10920, 2023. 

[13] Y. Qiao, D. Su, H. Kong, S. Sukkarieh, S. Lomax, and C. Clark, “Individual cattle identification using a 

deep learning based framework,” IFAC-PapersOnLine, vol. 52, no. 30, pp. 318–323, 2019. 

[14] M. Williams, N. Mac Parthal´ain, P. Brewer, W. James, and M. Rose, “A novel behavioral model of the 

pasture-based dairy cow from gps data using data mining and machine learning techniques,” Journal of dairy 

science, vol. 99, no. 3, pp. 2063–2075,  2016. 

[15] Y. Li, H. Shu, J. Bindelle, B. Xu, W. Zhang, Z. Jin, L. Guo, and W. Wang, “Classification and analysis of 

multiple cattle unitary behaviors and movements based on machine learning methods,” Animals, vol. 12, no. 

9, p. 1060, 2022. 

[16] Y. Qiao, M. Truman, and S. Sukkarieh, “Cattle segmentation and contour extraction based on mask r-cnn 

for precision livestock farming,” Computers and Electronics in Agriculture, vol. 165, p. 104958, 2019. 

[17] K. Liakos, S. P. Moustakidis, G. Tsiotra, T. Bartzanas, D. Bochtis, and C. Parisses, “Machine learning based 

computational analysis method for cattle lameness predic- tion.” HAICTA, vol. 128, p. 139, 2017. 

[18] D. Wu, Q. Wu, X. Yin, B. Jiang, H. Wang, D. He, and H. Song, “Lameness detection of dairy cows based 

on the yolov3 deep learning algorithm and a relative step size characteristic vector,” Biosystems Engineering, 

vol. 189, pp. 150–163, 2020. 

[19] S. Grampurohit and C. Sagarnal, “Disease prediction using machine learning algorithms,” in 2020 

International Conference for Emerging Technology (INCET). IEEE, 2020, pp. 1–7. 

[20] N. V. Kishan, Y. S. Trinath, S. Kavalur, V. Sangamesha, and S. Reddy, “Cattle disease identification using 

prediction techniques,” IJARIIE-ISSN (O), pp. 2395– 4396, 2021. 

[21] R. Dulal, L. Zheng, M. A. Kabir, S. McGrath, J. Medway, D. Swain, and W. Swain, “Automatic cattle 

identification using yolov5 and mosaic augmentation: A com- parative analysis,” in 2022 International 

Conference on Digital Image Computing: Techniques and Applications (DICTA).  IEEE, 2022, pp. 1–8. 

[22] N. Abdul Ghafoor and B. Sitkowska, “A machine learning application to predict risk of mastitis in cattle 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 31 No. 2 (2024) 

 

 

388 
https://internationalpubls.com 

from ams sensor data,” AgriEngineering, vol. 3, no. 3, pp. 575–583, 2021. 

[23] I. Ramirez-Morales, L. Aguilar, E. Fernandez-Blanco, D. Rivero, J. Perez, and A. Pazos, “Detection of 

bovine mastitis in raw milk, using a low-cost nir spectrom- eter and k-nn algorithm,” Applied Sciences, vol. 

11, no. 22, p. 10751, 2021. 

[24] D. M. Amin, G. Shehab, R. Emran, R. T. Hassanien, G. N. Alagmy, N. M. Hagag, M. I. Abd-El-Moniem, 

A. R. Habashi, E. M. Ibraheem, and M. A. Shahein, “Diagnosis of naturally occurring lumpy skin disease 

virus infection in cattle using virological, molecular, and immunohistopathological assays,” Veterinary 

World, vol. 14, no. 8, p. 2230, 2021. 

[25] G. Limon, A. A. Gamawa, A. I. Ahmed, N. A. Lyons, and P. M. Beard, “Epidemi- ological characteristics 

and economic impact of lumpy skin disease, sheeppox and goatpox among subsistence farmers in northeast 

nigeria,” Frontiers in veterinary science, vol. 7, p. 8, 2020. 

[26] L. Aerts, A. Haegeman, I. De Leeuw, W. Philips, W. Van Campe, I. Behaeghel, L. Mostin, and K. De Clercq, 

“Detection of clinical and subclinical lumpy skin disease using ear notch testing and skin biopsies,” 

Microorganisms, vol. 9, no. 10, p. 2171, 2021. 

[27] V. Punyapornwithaya, O. Arjkumpa, N. Buamithup, C. Jainonthee, R. Salvador, and K. Jampachaisri, “The 

impact of mass vaccination policy and control measures on lumpy skin disease cases in thailand: insights 

from a bayesian structural time series analysis,” Frontiers in Veterinary Science, vol. 10, p. 1301546, 2024. 

[28] N. Murti, E. Safi P. Srianto, S. Madyawati, and A. Rofikoh, “Prevalence and progression of lumpy skin 

disease cases in cattle over the six months leading up to eid al-adha in 2023 in the cirebon district of west 

java province, indonesia,” in IOP Conference Series: Earth and Environmental Science, vol. 1292, no. 1. 

IOP Publishing, 2024, p. 012037. 

[29] G. Mackereth, K. Rayner, A. Larkins, D. Morrell, E. Pierce, and P. Letchford, “Surveillance for lumpy 

skin disease and foot and mouth disease in the kimberley, western australia,” Australian Veterinary 

Journal, 2024. 

[30] A. Elsonbaty, M. Alharbi, A. El-Mesady, and W. Adel, “Dynamical analysis of a novel discrete fractional 

lumpy skin disease model,” Partial Differential Equations in Applied Mathematics, vol. 9, p. 100604, 2024. 

[31] Alaria, S. K. "A.. Raj, V. Sharma, and V. Kumar.“Simulation and Analysis of Hand Gesture Recognition 

for Indian Sign Language Using CNN”." International Journal on Recent and Innovation Trends in 

Computing and Communication 10, no. 4 (2022): 10-14. 

[32] Ashish Raj, Vijay Kumar, S. K. A. V. S. Design Simulation and Assessment of Prediction of Mortality in 

Intensive Care Unit Using Intelligent Algorithms. MSEA 2022, 71, 355-367.