http://drr.vau.ac.lk/handle/123456789/251 2026-07-28T20:35:12Z 2026-07-28T20:35:12Z Thapalingasivam, K. Vinoharan, V. Sivanesan, P. http://drr.vau.ac.lk/handle/123456789/2154 2026-07-27T03:47:11Z 2024-01-01T00:00:00Z

Classification of Sri Lankan Skin Diseases Using Ensemble Deep Learning Models Thapalingasivam, K.; Vinoharan, V.; Sivanesan, P. Skin disease detection is a critical component of dermatological healthcare, traditionally dependent on dermatologists’ expertise for accurate diagnosis. However, manual diagnosis can be time-consuming and prone to errors, especially in regions with limited access to specialized care. To address these issues, this study proposes a novel approach utilizing deep learning techniques for the automated detection of skin diseases. We employed state-of-the-art deep learning architectures, including VGG16, VGG19, EfficientNetB0, and ResNet50, to classify four prevalent skin conditions: Eczema (Dermatitis), Psoriasis, Tinea, and Vitiligo. A comprehensive dataset of 600 images from these classes was collected from Vavuniya General Hospital, Sri Lanka. To our knowledge, this is the first study to apply deep learning for skin disease detection specifically South Asian skin types, particularly type IV and V in Sri Lankan patients. The images underwent preprocessing, annotation, and data augmentation to enhance the models’ ability to capture distinct features of each condition. Performance evaluations revealed that VGG16 and ResNet50 achieved accuracies of 87%, while VGG19 and EfficientNetB0 also showed strong results. To further improve predictive performance, these models were combined into an ensemble model, achieving a final accuracy of 91%. To make this research practically applicable, a Flask application was developed that allows users to upload an image of a skin disease that predicts the disease name. This research fills a significant gap in medical image analysis, providing a foundation for future advancements in automated skin disease detection.

2024-01-01T00:00:00Z Vamadevan, S. Vinoharan, A. Vinoharan, V. http://drr.vau.ac.lk/handle/123456789/2153 2026-07-27T03:43:17Z 2026-01-01T00:00:00Z

Deep-Machine Learning Integration for Enhanced Pest Detection in Mangifera indica (Mango) Vamadevan, S.; Vinoharan, A.; Vinoharan, V. Mango (Mangifera indica), the third most widely consumed fruit in Sri Lanka after banana and papaya, has experienced steady growth in cultivation and yield; however, pest and disease infestations still cause substantial yield losses and degrade fruit quality. Since most pests inhabit mango leaves, early and accurate leaf-based detection is crucial for effective crop management, yet traditional manual inspection remains time consuming, labor-intensive, and often unreliable due to pest evolution and limited expert availability. To address these challenges, this study proposes an automated pest detection framework that integrates deep learning–based feature extraction with machine learning classification. The research uses ten classes from the MangoPestClassification dataset, consisting of real-world images of healthy and pest-affected mango leaves captured under diverse environmental conditions to represent realistic field scenarios. Pre-trained convolutional neural network (CNN) models, VGG19, ResNet50, InceptionV3, DenseNet121, MobileNetV2, and UNet, are employed to extract discriminative image features. These features are then classified using Logistic Regression, Random Forest, and Support Vector Machine (SVM) classifiers. This hybrid deep-learning–machine-learning approach ensures both strong feature representation and efficient classification. A 10-fold cross-validation strategy is applied to ensure reliable performance evaluation. Experimental results show high accuracy and robust ness across model combinations. ResNet50 features with Logistic Regression achieved 96.65±1.06% accuracy using 200 images per class, while ResNet50 features with an SVM (RBF kernel) yielded 99.09±0.30% accuracy with 1,000 images per class. Overall, the findings highlight that ResNet-based feature extraction combined with either SVM (RBF) or Logistic Regression offers an effective and practical solution for automated mango pest detection, reducing dependence on manual monitoring.

2026-01-01T00:00:00Z Vamadevan, S. Vinoharan, A. Vinoharan, V. http://drr.vau.ac.lk/handle/123456789/2152 2026-07-27T03:38:34Z 2025-01-01T00:00:00Z

Machine Learning for Effective Pest Detection in Mangifera indica (Mango) Crop Protection Vamadevan, S.; Vinoharan, A.; Vinoharan, V. Mangifera indica, a key fruit crop in Sri Lanka, supports both agriculture and rural livelihoods but suffers significant losses from pest infestations. Early pest detection is crucial to reduce damage, yet manual inspection is laborious and inefficient for large farms. Advances in artificial intelligence enable automated pest detection, offering a more accurate and efficient approach to mango crop protection. This study presents a machine learning approach for automatically detecting pests on mango leaves, providing farmers and agricultural experts with a practical tool for timely intervention. The study uses ten classes from the MangoPestClassification dataset, with 1,000 images per class, comprising real-world images of healthy and pest-affected mango leaves captured under diverse environmental conditions, thereby representing realistic field scenarios. To prepare themango leaf images for classification, three feature extraction techniques were applied: Histogram of Oriented Gradients (HOG) to capture shape and edge information, Bag of Features (BoF) to represent local descriptors, and Wavelet transforms to analyze texture and frequency components. These extracted features were then used to train and evaluate three machine learning classifiers: Support Vector Machines (SVM), Random Forest, and Logistic Regression. To ensure robust and reliable performance, 10-fold cross-validation was employed, providing an accurate assessment of the models’ classification capabilities. The experimental results indicated that with 200 images per class, the best performance was achieved using HOG+Wavelet with Random Forest at an accuracy of 81.20%, while with 1000 images per class, HOG+PCA with SVM (RBF) reached 91.64%. Overall, HOG-based features combined with SVM (RBF) or Random Forest provide the most reliable framework for mango pest classification and establish a strong foundation for practical automated detection systems. The study highlights the potential of machine learning to enhance pest monitoring in mango cultivation. By minimizing reliance on manual inspections and enabling early pest detection, the proposed method can contribute to increased yields, better fruit quality, and more sustainable mango farming practices in Sri Lanka.

2025-01-01T00:00:00Z Herath, W. R. G. A. D. P. Vinoharan, V. Kumaralingam, L. Nagulan, R. Bramya, S. http://drr.vau.ac.lk/handle/123456789/2151 2026-07-28T09:36:19Z 2024-01-01T00:00:00Z

Adapting Prompt-Driven Approach for Diabetic Foot Ulcer Segmentation: A Sri Lankan Patient Study Herath, W. R. G. A. D. P.; Vinoharan, V.; Kumaralingam, L.; Nagulan, R.; Bramya, S. Diabetes, a widespread metabolic disorder affecting millions globally, including many in Sri Lanka, often results in diabetic foot ulcers (DFUs). Accurate wound measurement is vital for diagnosing and monitoring DFUs, providing key metrics to track healing and inform treatment decisions. However, manual measurement is both time-consuming and prone to inaccuracies, often leading to misdiagnoses, and flawed documentation. Automated wound segmentation from medical images presents a promising solution by automating wound area assessment and enhancing patient outcomes. This study proposes a prompt driven segmentation approach to tackle the unique challenges of DFU segmentation, particularly for a limited dataset of Sri Lankan patients. This approach enables the model to use user input like points, regions, or bounding boxes to focus on complex areas, enhancing accuracy in cases with unclear wound boundaries or sparse data. The primary objectives of this research include the creation of a comprehensive dataset reflecting diverse DFU instances in Sri Lanka and the adaptation of the Segment Anything Model (SAM) to accurately segment DFU areas with minimal supervision. The dataset consists of 300 samples from 180 patients at Dambulla Base Hospital, capturing variations in skin tone, ulcer size, shape, and surrounding tissue conditions. The pre-trained SAM was fine-tuned using prompt driven segmentation techniques to handle the unique characteristics of Sri Lankan patients. The model demonstrated strong performance, achieving a Dice Coefficient of 83.31%, Specificity of 99.73%, and Precision of 90.42%. This approach minimizes the need for large annotated datasets, enhances generalisation from limited data, and allows user-guided refinement. The development of population-specific models in this study has the potential to enhance clinical decision-making and significantly improve patient care for DFU management in Sri Lanka.

2024-01-01T00:00:00Z