Research Article

A Deep Learning Framework for Automated Lumpy Skin Disease Detection in Cattle

Ithaya Rani Panneer Selvam1*, Sanasam Chanu Inunganbi2, K.B. Anusha3, G. Balakrishna4, Mohebbanaaz5, D. Sandhya Rani6

1Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, India; 2Department of Computer Science and Engineering, IIIT Manipur, India; 3Department of CSE–AIML & IOT, Vallurupalli Nageswara Rao Vignana Jyothi Institute of Engineering & Technology, Hyderabad, Telengana; 4Department of EEE, Srinivasa Ramanujan Institute of Technology, Anantapur, India; 5Department of ECE, Ravindra College of Engineering for Women, Pasupula Village, Nandikotkur Road, Kurnool- 518452 Andhra Pradesh, India; 6Department of Computer Science and Informatics, Mahatma Gandhi University, Nalgonda, Telangana, India.

Received | March 08, 2026; Accepted | April 05, 2026; Published | June 13, 2026

*Correspondence | Ithaya Rani Panneer Selvam, Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, Andhra Pradesh, India; Email: [email protected]

Citation | Selvam IRP, Inunganbi SC, Anusha KB, Balakrishna G, Mohebbanaaz, Rani DS (2026). A deep learning framework for automated lumpy skin disease detection in cattle. Adv. Anim. Vet. Sci., 14(6):1231-1244.

DOI | https://dx.doi.org/10.17582/journal.aavs/2026/14.6.1231.1244

ISSN (Online) | 2307-8316

Copyright: 2026 by the authors. Licensee ResearchersLinks Ltd, England, UK.

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Introduction

Food systems and rural economies in many areas of the world have been built on cattle production, providing a significant contribution to nutrition, employment, and source of income. The Lumpy Skin Disease (LSD) is a highly contagious viral disease that significantly impacts livestock productivity through reduced milk yield, skin damage, and trade restrictions. These restrictions often delay intervention, allowing the disease to spread further. This underlines the importance of developing reliable, affordable, and automated diagnostic tools. Such diseases should be identified correctly and within the required time to contain and treat them. Nevertheless, the practices that are in use in the diagnosis are reliant on visual examination, laboratory testing and expert veterinary assessment. Such methods are also expensive, need special infrastructure and can be impractical in isolated or underdeveloped areas. Thus, the time lag in the diagnosis might lead to uncontrolled infection and mortality of livestock.

Recent advances in deep learning have created opportunities to detect the diseases in the image data with the help of automated methods. Convolutional Neural Networks (CNNs) are one of the techniques that have received specific attention because of their ability to automatically detect meaningful features in images without the intervention of a manual person. CNNs have become particularly popular because they are capable of automatically acquiring hierarchical visual representations and directly out of the image data. Several studies have demonstrated promising results using pretrained and fine-tuned CNN models for LSD classification. For example, Abdullah et al. (2025) and Shakeel et al. (2024) reported high classification accuracy using transfer learning techniques, whereas Olorunshola et al. (2025) has proposed a modified CNN architecture, which had a higher detection rate due to the use of optimized features learning.

Despite these advancements, existing approaches exhibit important limitations. Many studies rely on a single CNN architecture without systematic comparison, making it difficult to determine the most robust model. Additionally, several methods depend on limited or curated datasets, raising concerns about their generalization to real-world conditions. Most existing works focus primarily on classification accuracy while neglecting interpretability and feature-level understanding, which are critical for practical veterinary adoption.

Therefore, a key research gap remains in developing a unified framework that (i) systematically compares multiple state-of-the-art CNN architectures, (ii) enhances feature representation for complex lesion patterns, and (iii) ensures both generalization and interpretability. To address this gap, the present study proposes an automated image-based classification system for distinguishing healthy and LSD-infected cattle. The approach evaluates multiple CNN architectures, including DenseNet121, InceptionV3, VGG16, VGG19, and Xception, within a consistent experimental framework. In addition, a Histogram Patch-Based Gabor Wavelet Feature Extraction method is integrated to improve texture sensitivity and multi-scale feature representation.

The main contributions of this work are:

• A systematic comparative analysis of multiple CNN architectures for LSD detection
• Integration of histogram enhancement and patch-based Gabor wavelet features with deep learning
• Improved model generalization and interpretability for real-world deployment

Section 2 is the literature review; Section 3 proposes a technique; Section 4 reports findings and observations; and the conclusion and future work are presented in Section 5.

Literature survey

Automated disease diagnosis in livestock has gained significant attention due to the inherent limitations of traditional veterinary diagnostic methods, which are often time-consuming, subjective, and dependent on expert knowledge. With the rapid advancement of artificial intelligence, deep learning particularly Convolutional Neural Networks (CNNs) has become a dominant approach for image-based cattle disease classification. However, the effectiveness of different CNN architectures varies depending on the visual characteristics of Lumpy Skin Disease (LSD), which typically presents as irregular lesions with diverse sizes, textures, and non-uniform color distributions. These characteristics require models capable of capturing both fine-grained texture details and multi-scale spatial patterns. Despite promising results, existing studies reveal inconsistencies in model selection, feature extraction strategies, and real-world applicability, indicating the need for more comprehensive and hybrid frameworks.

A substantial body of research focuses on transfer learning using pretrained CNN architectures. Abdullah et al. (2025) evaluated models such as DenseNet121, InceptionV3, MobileNetV2, and Xception for Lumpy Skin Disease (LSD) detection and reported superior performance of DenseNet121 due to its efficient feature reuse mechanism. Dense architectures such as DenseNet121 are particularly suitable for LSD detection because their dense connectivity preserves low-level texture features while enabling deeper semantic understanding, which is essential for identifying subtle lesion patterns. While deep architectures improve classification accuracy, their high computational complexity limits their deployment in resource-constrained environments. This highlights a key trade-off between accuracy and efficiency that remains unresolved in current literature.

To improve model robustness, Shakeel et al. (2024) employed data augmentation techniques, enhancing performance under varying illumination and background conditions. Although effective, such approaches depend on artificially generated variations and may not fully represent real-world complexities, thereby limiting generalization. Similarly, comparative studies by Saha (2024) and Senthilkumar et al. (2024) emphasized the importance of lightweight architectures like MobileNetV2, which offer reduced computational cost and suitability for real-time applications. However, the effectiveness of lightweight architectures reported by Saha (2024) is context-dependent, as these models prioritize computational efficiency over deep feature representation. While they perform well in real-time scenarios, they may not be optimal for LSD detection, where lesions exhibit complex textures, irregular shapes, and multi-scale variations. In contrast, Abdullah et al. (2025) demonstrated that deeper architectures such as DenseNet121 achieve higher accuracy by effectively capturing these fine-grained and multi-scale lesion characteristics. Therefore, while lightweight models are advantageous for efficiency, deeper architectures are more suitable when classification accuracy is the primary objective. Hybrid feature extraction approaches have been explored to enhance deep learning models.

In the case of LSD detection, Reza et al. (2026) produced an image classification system that utilized cattle mouth and hoof images based on a CNN system. Their findings demonstrated that deep learning models could be used effectively in discriminating between infected and healthy cows in the absence of manual feature extraction. Olorunshola et al. (2025) improved the LSD classification by adding CNN features to Haralick texture descriptors, and results were better than those of CNN-only models. Building on this, multi-stage feature extraction methods further improve performance by combining histogram-based enhancement for contrast improvement with patch-based Gabor wavelet features for capturing localized texture and frequency information. This integration enables more effective feature representation by combining global intensity patterns with detailed texture characteristics of LSD lesions.

In addition to recent advancements, foundational studies from 2015–2020 have established the effectiveness of Gabor wavelets in image analysis. For instance, Jain et al. (1991), Hosny et al. (2017), and Arivazhagan et al. (2026) demonstrated that Gabor filters effectively capture texture, orientation, and frequency information for classification tasks. These foundational works provide a strong theoretical basis for applying Gabor wavelet features to detect texture-rich skin lesions such as LSD. Despite its potential, such integrated frameworks remain underexplored in existing literature. Ensemble and hybrid learning strategies have also been investigated to improve robustness and generalization. Khandait et al. (2025) demonstrated that combining multiple pretrained CNN models enhances classification performance across diverse datasets. Similarly, Goddeti et al. (2024) integrated CNN feature extraction with Extreme Learning Machines (ELM), achieving faster training and high diagnostic accuracy. While these methods improve reliability, their increased architectural complexity can hinder scalability and real-time deployment.

Recent research points toward ensemble deep learning and object detection models of Lumpy Skin Disease (LSD) detection. Saqib et al. (2024) showed that MobileNetV2 provides a balance between accuracy and computational cost for field applications, while Sentoso et al. (2025) reported that VGG16 performs well on moderate-sized datasets, indicating that performance depends on dataset characteristics. Architectures such as Inception and Xception are particularly suitable for LSD detection due to their ability to capture multi-scale lesion patterns. However, these findings reveal no clear consensus on the optimal model architecture. Advanced approaches, such as the ensemble CNN proposed by Ayon et al. (2026) and the hybrid YOLO-EfficientNet model (LUMPNet) by Ubaidullah et al. (2026), achieve high accuracy but focus on specific architectures and lack a unified framework integrating both handcrafted and deep features. Furthermore, the broader literature indicates that CNN-based approaches, particularly when combined with transfer learning and data augmentation techniques, are highly effective for cattle disease diagnosis. However, key research gaps remain:

• Limited integration of histogram enhancement and patch-based texture feature extraction (e.g., Gabor wavelets) with deep CNN models.
• Lack of a systematic comparison of multiple CNN architectures within a unified framework.
• Insufficient focus on balancing accuracy, computational efficiency, and real-world applicability.

To address these limitations, and considering the texture-rich and multi-scale nature of LSD lesions, the present study proposes a Histogram Patch-Based Gabor Wavelet Feature Extraction integrated with Deep CNN framework. This approach is designed to complement CNN architectures by enhancing texture sensitivity and multi-scale representation, thereby improving classification performance.

Block diagram

Figure 1 illustrates the proposed deep learning–based image classification system that transforms raw cattle images into meaningful class labels. The workflow begins with image acquisition, followed by resizing images to a uniform dimension and pre-processing using Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance image contrast and highlight disease-affected regions. The enhanced images are then divided into non-overlapping patches to capture localized information. Gabor wavelet filters are applied to each patch to extract texture-related features, such as lesion patterns and skin irregularities, by capturing spatial frequency and orientation characteristics. The features extracted from all patches are subsequently combined into a unified feature vector. This feature representation is then provided to multiple pre-trained deep learning architectures, including CNN, DenseNet121, ResNet50V2, InceptionV3, VGG16, VGG19, and Xception, for feature learning and classification. The final output categorizes cattle images into healthy or LSD-infected classes.

CNNs are deep learning models which identify hierarchical patterns in images for tasks like detection and classification. Their main components convolutional, pooling, flattening, and fully connected layers work together: convolutional layers extract features from low to high levels, pooling layers reduce computation, and fully connected layers perform classification and depth wise separable convolutions. Advanced CNN architectures, such as ResNet50V2, InceptionV3, VGG16, VGG19, and Xception, enhance performance through techniques like residual connections, extraction of feature at multi-scale, convolutions which are separable on the basis of depth. ResNet50V2 improves gradient flow, InceptionV3 captures multi-scale features, VGG16/19 use simple filters with deep extraction, and Xception reduces computation while preserving rich features, making these models effective for complex image classification tasks.

 

Materials and Methods

Pre-processing using histogram techniques

Pre-processing is essential to improve image quality and optimize model performance. Raw images may contain noise, low contrast, and uneven illumination, which can reduce feature extraction efficiency. To address these issues, adaptive histogram techniques such as CLAHE are applied, which redistribute the intensity values of pixels locally to enhance contrast across different regions of the image. Unlike standard global histogram equalization, CLAHE operates on small tiles of the image, ensuring that regions with varying illumination are enhanced without over-amplifying noise, thereby improving the visibility of skin lesions. The transformation can be expressed as:

....(1)

Where L is the number of gray levels, nj is the number of pixels at intensity j, and N is the total number of pixels. In our framework, this adaptive enhancement ensures that lesion features are consistently highlighted across all regions, improving the efficiency of subsequent Gabor feature extraction and CNN-based classification an shown in Figure 2.

Texture enhancement using patch wise gabor wavelet filtering

After histogram-based enhancement, the improved cattle images S are subjected to Gabor filtering (Hosny et al 2017) for texture analysis. Each enhanced cattle image is convolved with a set of Gabor kernels at different orientations and wavelengths. This operation enhances directional textures such as wrinkles, spots, and lesion boundaries on the animal’s skin. The resulting output, commonly referred to as the Gabor magnitude response, emphasizes high-frequency components such as edges and irregular patterns while suppressing uniform regions. Mathematically, the filtered output can be expressed as:

I ⊆ S…(2)

Each enhanced cattle image I is divided into non-overlapping patches which is represented as P (P= 1, 2, ….., (m×n)). The m and n are number of patch arranged in row and column respectively. Total number of patches m, n. Typically, the m1 and n1 are the row and column of each patch size (i.e pixels) respectively.

Ip(a,b)+I(a:a+m1,b:b+n1)......3

Where; a=1: r1 step by m1; b=1:c1 step by n1

Where; r1 and c1 are numbers of rows and columns in the cattle image (256 x 256 pixels ) respectively, m1 and n1 are the row and column of each patch size, respectively.

Ig,p (x, y) = Ip (x, y) * g (x, y)......4

Where; Ip (x, y) represents the enhanced each patch cattle image and g(x, y) is the 2D Gabor Kernel. The symbol * denotes convolution. The magnitude is then computed as:

…(5)

Where g (x, y) is the Gabor filter response, σ control scale, f is frequency and θ is orientation, IG capturing texture and directional feature of LSD. Features extracted from all patches are concatenated into a single feature vector as IG. This process yields a feature vector of size (number of patches × number of scales × number of orientations). This magnitude image provides a clearer representation of texture variations, making it highly effective for identifying abnormal skin patterns in cattle as shown in Figure 2.

 

Multimodal architecture

Convolution neural network (CNN)

A Convolutional Neural Network (CNN) is a deep learning model used for image analysis, as proposed by Saha (2024). It automatically learns spatial hierarchies of features from the input through convolutional layers, which use filters to extract patterns such as edges, textures, and shapes. These features are down sampled using pooling layers to reduce computational complexity and prevent over fitting. For final classification the mapping of features that are extracted are done for flattening as well as passing it to layers that are fully connected. CNNs use activation functions like ReLU to introduce non-linearity and softmax for binary-class prediction. Due to weight sharing and local connectivity, CNNs efficiently manage large image data and are widely applied in medical imaging, object detection, and disease diagnosis. Hence, CNN are used for cow disease classification, with improved input quality, training on diverse data, and evaluation on a separate test set to ensure generalization and prevent overfitting, as defined in the corresponding Equation 2.

For input layer:

X ⊆ IG

Convolutional Layer 1:

H1 = Relu (W1 * X + b1)....6

Where; W1 and b1 represents weights and biases for first convolutional layer which represent convolution operation. H1 represents output feature map after applying 32 filter of size 3 x3. The convolution neural network was implemented using the Keras Sequential API. It consisted of three consecutive convolution networks with the number of filters slowly varying (32, 64, 128) and 2× 2 max-pooling operations that followed to extract discriminative features and reduce spatial dimensions. The resulting feature maps were flattened and passed to a fully connected layer with 128 neurons activated by ReLU. To reduce over-fitting, a dropout rate was applied before the output layer. The last layer utilized the softmax activation function to produce class probability distributions among the two categories, namely, healthy and lumpy skin disease.

Max Pooling Layer 1:

P1= MaxPool (H1, Pool_size = 2)....7

In this case, P1 reduces the spatial resolution of H1 by 2. The Adam algorithm was used to optimize the network where the loss function was categorical cross-entropy. The training was done on 20 epochs and the classification accuracy was monitored during that period to evaluate the convergence behaviour and possible overfitting.

Convolutional Layer 2:

H2 =ReLu (W2*P1+b2) v.....(8)

DenseNet121

DenseNet121 is a deep convolutional architecture proposed by Huang et al. (2017) that contains 121 layers organized through densely connected blocks. In this structure, each layer obtains inputs from all previous layers within the same block, enabling continuous feature reuse and improved information flow. This connectivity pattern strengthens gradient propagation, mitigates vanishing gradient issues, and reduces redundant parameters compared to conventional CNN models. The conventional convolution and pooling operations are performed using network made up of number of dense blocks, followed by a final classification stage. Owing to its compact design and strong feature extraction capability, DenseNet121 is extensively applied in image analysis and medical diagnosis tasks. The input of the model is some batch of images in the form of tensors (B,H,W) when B is the size of the batch, and H and W are the image height and width, respectively.

DenseNet121 Backbone can be stated as:

Y1 =F1([X0, X1,..,Xn-1]).......(9)

In the expression, X0, X1, .., Xn-1 denote the feature maps produced by all preceding layers, and the bracket notation [.] represents channel-wise concatenation of these feature maps, rather than addition. The function Fl (.) denotes a composite transformation consisting of batch normalization, ReLU activation, and convolution. This formulation highlights the dense connectivity pattern, where each layer receives inputs from all previous layers, promoting feature reuse and improved gradient flow. These metrics reveal both strengths and limitations of the model.

Global average pooling

For each channel c:

Vector form:

g = [g1, g2,…,gc]

Where H and W represents spatial dimensions of the last feature map, and c is the channel index.

ResNet50V2

ResNet50V2 is a 50-layer residual network introduced by Goceri (2019). It employs identity shortcut connections that bypass intermediate layers, enabling improved gradient flow and preventing degradation in deep networks. The V2 variant adopts a pre-activation structure, where batch normalization and activation precede convolution, enhancing training stability. Owing to its strong feature extraction and efficient optimization, ResNet50V2 is widely used in image classification and medical imaging tasks.

Feature extraction with ResNet50V2 (pre-activation )

F= x+W3.σ (BN (W2. σ (BN (W1. σ (BN (x))))))...(11)

Where; x is the input feature map from the previous layer. BN is the batch normalization, σ is the ReLU activation non-linearity. W is the activation precede convolution filter. ResNet50V2 was chosen because its deep residual design uses skip connections to allow gradients to propagate more effectively through many layers, avoiding vanishing gradient issues common in deep CNNs. Transfer learning was applied by starting network with ImageNet-trained weights, letting it reuse general visual features. After freezing the base convolutional layers, a new classification head with global average pooling, a dense ReLU layer, dropout (0.5), and a final softmax output for two classes was included.

Global average pooling

Global Average Pooling (GAP) is applied after the convolution base of the ResNet50V2 to reduce the spatial dimensions of feature maps while preserving important information. Unlike fully connected layers, GAP significantly reduces the number of parameters and helps prevent overfitting.

…(12)

Where; h and w stand for the height and breath of the feature map. Fi,j,k is the activation at position (I, j) in the kth feature map

Fully connected layers:

First dense layer is given by:

h1=Relu(W1z+b1)

Dropout Layer (Regularization):

h1dropout =Dropout (h1, rate=0.5)....(13)

Output Layer (Softmax Activation) is given by:

Where; C is the number of classes (2 in our case: healthy and lumpy skin), and y ̂l represents the predicted probability for class j.

InceptionV3

InceptionV3 is proposed by Jyotsna et al. (2018). The model utilizes inception modules that perform parallel convolutions with different filter sizes to extract multi-scale features efficiently. It also incorporates factorized convolutions, batch normalization, and auxiliary classifiers to reduce computational cost and enhance training convergence. Owing to its optimized structure and strong feature learning capability, this preparation helps the model learn distinct features and reduces overfitting for more accurate classification.

Standard convolution

Equation (15)

Computational cost can be given as:

Cost5×5= 25MN

Cost3×3+3×3= 18MN

Where, M is the input channel and N is output channel.

(b) Asymmetric factorization:

5 × 5 ≈ 3 × 3 + 3 × 3

1×1 Convolution (Dimensionality Reduction). (Equation (16)

Used for reducing feature depth before expensive convolutions.

Softmax classification:

Equation (17)

Where, zc is the logit (raw output) for class c, C represents total number of classes, and zk represents logit for class k. The model was trained for 20 epochs with a batch size of 32 using the Adam optimizer and categorical cross-entropy loss, while accuracy was tracked as the evaluation metric. Validation data helped monitor generalization and detect overfitting during training. Afterward, the trained model was saved for later use on new images.

VGG16

VGGNet, introduced by Simonyan et al. (2015), is a deep CNN known for its uniform architecture. It stacks small 3×3 convolution layers with periodic max-pooling, followed by fully connected layers for classification. The straightforward design enables effective hierarchical feature extraction, making VGG widely used in image analysis tasks. Hence, The VGG16-based classifier is used to sort cow images into healthy and LSD.

Convolution operation (3×3 Kernel)

Where; Xm represents input feature map, Wk,m denotes 3×3 convolution kernel, bk represents bias, M represents number of input channels, and Yk denotes output feature map.

ReLU activation

f(x) = max(0, x)....(19)

Max Pooling (2×2)

Y(i,j)= max (u,v)∈Ω X(i+u,j+v)....(20)

Where Ω represents the 2×2 pooling region.

Softmax classifier

Where; C states the number of classes, and zc is logit for class c.

VGG19

The VGG19 is a deep convolutional neural network with 19 weight layers, developed by Simonyan et al. (2015). It is built using repeated 3×3 convolutional layers with ReLU activation, followed by periodic max-pooling for spatial reduction and fully connected layers for classification. The VGG19 model employs the same fundamental mathematical operations as VGG16, including convolution, activation, pooling, and softmax classification. Compared to VGG16, VGG19 greater depth enables learning more complex and hierarchical feature representations; however, it also increases the number of parameters, resulting in higher computational cost and a greater risk of overfitting, particularly when applied to limited datasets such as Lumpy Skin Disease (LSD) images.

Xception

An Xception, proposed by Chollet (2017), is a deep convolutional architecture that relies entirely on depth wise separable convolutions. This approach divides standard convolution into channel-wise spatial filtering followed by point wise convolution for feature integration, which reduces computational overhead while maintaining strong representational power. The network also incorporates residual connections to enhance gradient flow and training stability. Xception is an extensively applied in image classification tasks, including cattle disease detection.

ReLU activation

Xception is chosen because it uses depth wise separable convolutions to break standard convolutions into lightweight spatial and pointwise steps, reducing computation while still capturing detailed features for image classification. Transfer learning with ImageNet-pretrained weights lets the network reuse learned visual features for cow disease detection. On top of the frozen base, custom layers Global Average Pooling 2D, the ReLU-activated dense layer with dropout, and a softmax output produce probabilities for the two classes: healthy and lumpy skin.

Standard convolution

Equation (22)

Depth wise separable convolution (Xception)

(a) Depth wise convolution. Equation (23)

(b) Pointwise convolution (1×1) Equation (24)

Where, X states input feature map, K represents depth wise kernel, M states number of input channels and Y is the output feature map. Model evaluate ( ) is used to measure the trained network’s performance on previously unseen test samples by calculating the loss value and classification accuracy, thereby assessing its generalization capability. For further validation, randomly selected test images are visualized with their predicted categories and associated probability scores, which assists in spotting incorrect predictions. The saved model can also be loaded again to perform inference on new input images, enabling its effective application in cow disease diagnosis.

Experimental data set and result

An overview is provided in this section to introduce the findings and discussion of the proposed scheme. This study employs three publicly available datasets for experimental evaluation: The Mendeley LSD image dataset proposed by Kumar et al. (2022), the Kaggle Lumpy Skin Disease Images Dataset, and the Veterinary Research LSD dataset (collected from publicly available veterinary sources and prior published studies). These datasets are used to evaluate the performance of the proposed framework. Sample images from the three datasets are illustrated in Figure 3, and their detailed descriptions are presented in Table 1. For computational efficiency, the image resolutions in all datasets were uniformly resized to 256 × 256 pixels. The proposed research was conducted on a personal computer equipped with an Intel Core i5 processor running at 3.20 GHz with 8 GB of RAM. All experiments were implemented using MATLAB R2013a.

The models were trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Training was conducted for 50–100 epochs, with early stopping applied based on validation loss (patience= 10), using an 80:20 training–testing split. Overfitting was controlled through the use of dropout and data augmentation techniques, including rotation, flipping, and scaling. As shown in Table 1, class imbalance was addressed by applying data augmentation to the minority (infected) class, along with class weighting during training to balance the loss contribution. The use of these three datasets Mendeley, Kaggle, and the Veterinary Research LSD Dataset is significant for several reasons. They provide diversity in data sources, which helps reduce dataset bias and improves the reliability of the model.

The datasets include variations in image conditions such as resolution, lighting, pose, and background, thereby simulating real-world scenarios. Each dataset also exhibits differences in class distribution between healthy and infected cattle, enabling evaluation of model stability under imbalanced conditions. In addition, the combined datasets cover multiple cattle breeds and varying stages of Lumpy Skin Disease (LSD), enhancing the model’s ability to generalize across different cases. Furthermore, pre-processing and augmentation techniques, particularly in the Kaggle dataset, introduce variations such as spatial transformations, which help in assessing the robustness of the model.

Performance analysis of histogram techniques for LSD

In this experiment, raw cattle images were initially enhanced using the histogram-based technique CLAHE. The enhanced features were then provided as input to several pre-trained architectures, including DenseNet121, ResNet50V2, InceptionV3, Xception, a baseline CNN, VGG16, and VGG19. The models were independently evaluated on the Mendeley dataset, the Kaggle dataset, and the Veterinary research LSD dataset the results are presented in Figures 4, 5, and 6, respectively, for the classification of normal and infected skin. It is observed that integrating histogram-based enhancement consistently improves accuracy across all datasets compared to models trained without histogram technique.

 

Table 1: Overview of lumpy skin disease (LSD) image datasets and their characteristics.

Dataset name	Total images	Healthy	LSD - infected	Image size	Key Characteristics
Mendeley LSD Dataset (Kumar et al., 2022)	1,024	700	324	256 x 256	Includes multiple breeds (Santa Gertrudis, Brahman, Guzerat, Yak); diverse poses, lighting, and perspectives
Kaggle LSD Dataset	2,223	1,520	703	640 x 640	Kaggle LSD Dataset;larger dataset enhanced using augmentation (flipping, rotation, cropping, shear) for robustness
Veterinary research LSD Data Set	1,023	478	545	256 x 256	Real farm images with variations in breed, age, and disease severity

 

 

 

Performance analysis of gabor wavelet filtering for LSD

The model was evaluated using different numbers of image patches where a Gabor filter was applied to each patch to extract texture features. The extracted Gabor features from all patches were then concatenated to form a single feature vector, which was used for classification across three datasets: Mendeley, Kaggle, and the veterinary research LSD dataset. Experimental results indicate that using 64 patches arranged in 8 rows and 8 columns (8 × 8 grid) achieved the highest classification accuracy of 94.5%, 94%, and 95.8%, respectively, outperforming configurations with 16 and 256 patches. This demonstrates that the 64-patch configuration provides an optimal balance between capturing sufficient local texture details and maintaining computational efficiency, as presented in Table 2.

 

Table 2: Classification accuracy of LSD detection using different patch sizes.

Number of patches	Each patch size (Pixels)	Accuracy (%)
		Mendeley dataset	Kaggle dataset	Veterinary research LSD dataset
16 (arranged in 4 rows x 4 columns)	64x64	92.5	92.7	92
64 (arranged in 8 rows x 8 columns)	32x32	94.5	94	95.8
256 (arranged in 16 rows x 16 columns)	16 x 16	94	93	92.3

 

In this experiments, the orientation and scale parameters of the Gabor filter play a vital role in determining the effectiveness of texture-based feature extraction. A comparative evaluation was performed using different orientation sets, including 2 (0°, 90°), 4 (−45°, 0°, 45°, 90°), 6 (−45°, −22.5°, 0°, 22.5°, 45°, 90°), and 8 (−67.5°, −45°, −22.5°, 0°, 22.5°, 45°, 67.5°, 90°), while maintaining a fixed scale value of eight. Experimental results obtained across three datasets demonstrate that increasing the number of orientations beyond four leads to a decline in classification performance by approximately 2–3%. Conversely, the configuration with four orientations consistently outperforms the two-orientation setup, achieving nearly 2% higher accuracy across all datasets and image resolutions as shown in Table 3.

 

Table 3: Classification performance (%) obtained using four sets of orientation.

Number of orientation	Accuracy (%)
	Three data set
	Mendeley dataset	Kaggle dataset	Veterinary research LSD dataset
2	90	93	88
4	95	94	93
6	93	93	91
8	92	93	90

 

Table 4: Classification performance (%) obtained using four sets of scales.

Number of scales	Accuracy (%)
	Three data set
	Mendeley Data Set	Kaggle Data set	Veterinary research LSD dataset
4	91	94	92
6	93	95	94
8	95	96	96
10	94	95	95

 

Furthermore, the impact of varying scale parameters was analyzed by fixing the number of orientations at four and testing scale ranges of (5:2:11), (5:2:15), (5:2:19), and (5:2:23), corresponding to 4, 6, 8, and 10 scales, respectively. The findings indicate that reducing the number of scales from 8 to 6 and 4 results in a performance drop of about 2% and 4%. Although increasing the scales to 10 provides a marginal improvement of approximately 1% compared to 6 scales, the optimal balance between computational efficiency and classification accuracy is achieved with 4 orientations and 8 scales. This configuration proves to be the most effective for differentiating between healthy and infected skin samples across all three datasets as shown in Table 4. The output of the Gabor feature extraction produces a feature vector of size 32, derived from 8 scales and 4 orientations. This feature vector is then used as input to various CNN architectures, including DenseNet, ResNet50V2, InceptionV3, VGG16, VGG19, and Xception, for further classification and analysis.

Performance evaluation of several models on the veterinary research LSD dataset

The Veterinary Research LSD dataset includes 1,023 images (478 healthy, 545 infected), split 80:20 for training and testing. Performance evaluation of the models showed varied results, as shown in Table 5. Among them, the proposed Histogram + Patch-wise Gabor wavelet + CNN model achieved the best performance with an accuracy of 0.945, along with high F1-score (0.95), recall (0.96), and precision (0.95). Other models such as DenseNet121, ResNet50V2, InceptionV3, Xception, VGG16, and VGG19 produce comparatively lower accuracies, generally ranging between 83.15% and 88.66%, while maintaining F1-scores between 0.88–0.93. Overall, the findings indicate that the CNN integrated with histogram and patch wise Gabor features provides the most effective classification performance for LSD detection.

The classification results for this dataset are summarized in Table 5. Based on these results, the corresponding confusion matrix is presented in Table 6. Out of the 545 infected images, 523 were correctly identified (true positives), while 22 were misclassified as healthy (false negatives). Similarly, among the 478 healthy images, 450 were correctly classified (true negatives), and 28 were incorrectly predicted as infected (false positives). These results yield an overall accuracy of 0.945, precision of 0.95, recall of 0.96, and F1-score of 0.95, demonstrating strong performance of the proposed method despite the slight class imbalance in the dataset.

Performance comparison of state-of-the-art models for lumpy skin disease (LSD) detection

The performance comparison of LSD detection methods using an 80:20 training–testing split, demonstrates a clear improvement in model effectiveness as shown in Table 7. Earlier research by Evgin (2019), based on a ResNet architecture, reported accuracy values of about 80%, 83%, and 82.6%, respectively. Subsequent work by Swapnil et al. (2024) using a hybrid VGG16 + ANN model significantly improved results, achieving 92.5%, 91.6%, and 89% accuracy. Further advancements were observed in later studies, where Jyotsna et al. (2018) attained approximately 95%, 93%, and 94% using VGG16 and InceptionV3, while Saha (2024) reported accuracy levels between 86.4% and 90% with a CNN-based approach. Similarly, Shakeel et al. (2024) using Xception CNN and Abdullah et al. (2025) with DenseNet121 maintained consistent performance in the range of 90%–94% accuracy. More advanced techniques, including the Vision Transformer proposed by Ullah (2025) and the YOLO + EfficientNet (LUMPNet) model by Ubaidullah et al. (2026), achieved accuracy levels up to 0.95, while ensemble-based methods by Ayon et al. (2026) reported around 0.94 accuracy. In contrast, the proposed approach integrating Histogram features, Gabor filtering, and CNN achieves superior results, reaching an accuracy

 

Table 5: Performance of proposed method in veterinary research dataset for LSD.

Model Name	Accuracy (%)	F1 Score	Recall	Precision
Histogram + patch Gabor + DenseNet121	87.31	0.92	0.95	0.89
Histogram + Patch Gabor + ResNet50V2	83.15	0.91	0.95	0.88
Histogram + Patch Gabor + InceptionV3	83.55	0.89	0.94	0.84
Histogram + Patch Gabor + Xception	86.85	0.89	0.94	0.85
Histogram + Patch Gabor + VGG16	88.66	0.88	0.93	0.84
Histogram + Patch Gabor + VGG19	88.11	0.93	0.95	0.92
Histogram + Patch Gabor + CNN	94.52	0.95	0.96	0.95

 

Table 6: Confusion matrix for LSD detection on the veterinary research LSD dataset.

	Predicted healthy	Predicted infected
Actual healthy	450 (TN)	28 (FP)
Actual infected	22 (FN)	523 (TP)

 

of 96.5% on the Mendeley dataset, 96% on the Kaggle dataset, and 95% on the veterinary research LSD dataset, indicating its superior capability for accurate LSD classification.

Moreover, a ten-fold cross-validation strategy was adopted to ensure robust evaluation of the proposed model. The model achieved accuracies of 97.5%, 97.34%, and 98% on the Mendeley, Kaggle, and Veterinary Research LSD datasets, respectively. These results demonstrate the consistency and effectiveness of the proposed approach across multiple datasets, highlighting its strong generalization capability for Lumpy Skin Disease (LSD) detection.

Conclusion

The growing incidence of cattle diseases such as Lumpy Skin Disease underscores the need for sophisticated detection techniques. Conventional methods, including manual examination and laboratory analysis, are often costly, time-consuming, and prone to human error. The proposed framework, combining histogram-based pre-processing, patch-wise Gabor wavelet feature extraction, and deep learning models, offers a robust alternative. CLAHE enhances image contrast and highlights skin lesions. The enhanced images are then divided into non-overlapping patches, and texture features are extracted from each patch using Gabor wavelets. These features are combined into a single feature vector and fed into deep learning models, including CNN, DenseNet121, ResNet50V2, InceptionV3, VGG16, VGG19, and Xception, to classify cattle as healthy or infected. The proposed framework was evaluated on the Mendeley, Kaggle, and veterinary research LSD dataset, demonstrating superior performance compared to conventional deep learning methods. The results confirm that integrating image enhancement and Gabor-based texture analysis with deep learning achieves high detection accuracy, providing a fast, scalable, and cost-effective solution for early LSD diagnosis, thereby supporting timely veterinary intervention and informed livestock management.

 

Table 7: Comparison of existing LSD detection approaches with proposed methodology in three datasets.

Authors	Year	Model / Technique	Accuracy (%)
			Mendeley data set	Kaggle data set	veterinary research LSD dataset
Evgin Goceri et al.	2019	ResNet	80	83	82.6
Swapnil et al.	2024	VGG16 + ANN	92.5	91.6	89
Jyotsna et al.	2018	VGG16, InceptionV3	95	93	94
Saha et al.	2024	CNN	86.4	88	90
Muhammad Zain et al.	2024	Xception CNN	94.8	93.4	93
Walid Abdullah et al.	2025	DenseNet121	90.6	91	91
Ramakrishna et al.	2025	CNN	94	92.8	93
Asad Ullah et al.	2025	Vision Transformer (ViT)	94.12	93	92
Ubaidullah et al.	2026	YOLO + EfficientNet (LUMPNet)	94–95	95.3	94
Nazibul Basar Ayon et al.	2026	VGG16 + ResNet + Inception	94.2	93	94
Our Proposed (2026)	2026	Histogram + Patch GW + CNN	96.5	96	95

 

However, certain limitations exist, including dataset constraints, variations in image quality and background, and the absence of ensemble or hybrid feature extraction methods, which may affect generalizability. Future work will focus on incorporating real-time field datasets, standardizing image acquisition to ensure consistent capture conditions (lighting, angle, distance, and image quality), and exploring advanced modelling techniques to further enhance detection performance.

Acknowledgments

The authors would like to thank all the veterinarians, farm owners, and technical staff who assisted in data collection and provided valuable insights. We also acknowledge the support of Koneru Lakshmaiah Education Foundation for providing resources for this research.

Novelty Statement

A novel hybrid framework was introduced that integrated CLAHE-based image enhancement with patch-wise Gabor texture feature extraction to improve the visibility and representation of Lumpy Skin Disease lesions in cattle images. Furthermore, the proposed approach combined handcrafted texture analysis with deep learning classifiers, enabling superior detection performance across multiple benchmark datasets compared to existing state-of-the-art methods

Author’s Contribution

All authors participated in data analysis, manuscript drafting, and revisions, and consented to assume responsibility for all aspects of this study.

Generative AI and AI assisted technology statement

The authors declare that no generative AI and AI assisted technology was used in the creation of this manuscript.

Conflict of interest

The authors have declared no conflict of interest.

References

Abdullah W (2025). Transfer learning with DenseNet121 for automated livestock disease classification. IEEE Access, 13: 1–12.

Abdullah W, Tanwar S, Abouhawwash M (2025). Deep learning-based detection of lumpy skin disease in livestock using CNNs. Sustain. Machine Intell. J., 11: 1–10. https://doi.org/10.61356/SMIJ.2025.11515

Al-Zubi A (2024). Lumpy Skin disease detection in cattle by a robust approach using advanced convolutional neural networks. Indian J. Anim. Res., 58(12): 2146–2153. https://doi.org/10.18805/IJAR.BF-1793

Arivazhagan S, Ganesan L (2006). Texture classification using Gabor wavelets based rotation invariant features. Pattern Recogn. Lett., 27(16): 1976–1982. https://doi.org/10.1016/j.patrec.2006.05.008

Ayon NB, Rahman MS, Islam MR (2026). Ensemble deep learning approach for automated detection of lumpy skin disease and foot-and-mouth disease in cattle. Artif. Intell. Agric., 12: 2345-2353.

Chollet F (2017). Xception: Deep learning with depth wise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pp. 1251–1258. https://doi.org/10.1109/CVPR.2017.195

Evgin G (2019). Deep learning-based classification of skin lesions using ResNet architectures. Biomed. Signal Proc. Contr., 51: 1–8.

Goceri E (2019). Deep learning-based classification of skin lesions using ResNet architectures. Biomed. Signal Proc. Contr., 51: 1–8.

Goddeti S, Narayana V (2024). Lumpy skin disease diagnosis using deep feature extraction and extreme learning machines. Int.J. Eng.Trends and Tec, 72(9):106–113.

Hosny KM, Darwish M, Radwan I (2017). Texture classification using Gabor filters and neural networks. Int. J. Adv. Comp. Sci. Appl., 8(3): 123–130.

Huang G, Liu Z, Van Der Maaten L (2017). Densely connected convolutional neural networks. IEEE Conf. Comp.Vision and Pattern Recogn., 8(2): 4700–4708.

Jain AK, Farrokhnia F (1991). Unsupervised texture segmentation using Gabor filters. Pattern Recogn., 24(12): 1167–1186. https://doi.org/10.1016/0031-3203(91)90143-S

Jyotsna B, Nitin R (2018). Convolutional Neural Network based Inception v3 Model for Animal Classification, In. J. Adv. Res. Compu. Commu. Engg, 7(5): 45

Kaggle dataset link- publically available https://www.kaggle.com/datasets/shivamagarwal29/cow-lumpy-disease-dataset.

Khandait PD, Lanjewar K, Patle S, Meshram R (2025). Deep learning-based cattle disease detection: A CNN approach for identifying lumpy skin disease and foot-and-mouth disease. Int. J. Innov. Res. Technol., 11(10).

Kumar S, Shastri S (2022). Lumpy skin images dataset. Mendeley Data 2022, V1. Available: https://data.mendeley.com/datasets/w36hpf86j2/1.

Olorunshola E, Lakan N, Adamu F (2025). Hybrid convolutional neural network-haralick framework for foot-and-mouth disease classification in cattle. Asian J. Research in Computer Science, 18(6):390–402.

Reza MR, Sahriani P, Mukjizat S (2026). Clinical image-based classification of foot-and-mouth disease in cattle using convolutional neural networks. J. Sistem Teknol. Inf. Indonesia, 6(1): 22–31. https://doi.org/10.32528/justindo.v11i1.4737

Saha DK (2024). An extensive investigation of convolutional neural network architectures for diagnosing Lumpy Skin Disease in dairy cattle. Heliyon, 10(4): e27893. https://doi.org/10.1016/j.heliyon.2024.e34242

Saqib S, Khan M, Sharif A, Rehman A (2024). Lightweight CNN-based detection of Lumpy Skin Disease in cattle images. Computers in Biology and Medicine, 168:107692.

Senthilkumar C, Sindhu C, Vadivu G, Neethirajan S (2024). Early detection of lumpy skin disease in cattle using deep learning. A comparative analysis of pretrained models. Vet. Sci., 11(10): 510. https://doi.org/10.3390/vetsci11100510

Sentoso F, Ardiansyah R, Wangsa S (2025). Performance evaluation of VGG and ResNet architectures for cattle disease image classification. Sistemasi: J. Syst. Inf., 14(1): 89–98.

Shakeel M, Tauheed Z, Javaid N, Aslam T (2024). A deep learning tool for early detection and control of lumpy skin disease using convolutional neural networks. J. Comp. Biomed. Inf., 7(2): Article 547.

Simonyan K, Zisserman A (2015). Very deep convolutional networks for large-scale image recognition. Proc. Int. Conf. Learn. Represent. 82:1-14.

Swapnil S, Adit H, Sakshi S (2024). Cattle disease detection using VGG16 convolutional neural network architecture, Int.Conf. Comp. Commu. Net. Techn. 1- 6. doi: 10.1109/ICCCNT61001.2024.10724717.

Ubaidullah A, Hussain H, Jafri M (2026). LUMPNet: A hybrid YOLO–EfficientNet framework for lumpy skin disease detection. Comp. Biol. Med., 180: 107694.

Ullah A (2025). Vision transformer-based automated classification of cattle skin diseases. Expert. Syst. Appl., 235.