Conference papers

Permanent URI for this communityhttps://rda.sliit.lk/handle/123456789/4889

Browse

Filters

2025 - 20265

Advanced Search

Filter by

Settings

Subject: search.filters.subject.Deep Learning

Search Results

Now showing 1 - 5 of 5
  • Thumbnail Image
    ItemEmbargo
    Hybrid Deep Learning Approach Using YOLO and Spatial Transformers for Kinesthetic Math Education in Low-Resource Settings
    (Institute of Electrical and Electronics Engineers Inc., 2025-07-03) Sellapperuma M.S; Wasana, K.H. I; Dhananjana, B.K. T; Anuruddhika, D. L. N; Krishara, J; Wijendra, D
    This research presents a hybrid deep-learning system to enhance kinesthetic math education for Grade 4 and 5 students in low-resource Sri Lankan classrooms, addressing addition, subtraction, and time-telling skills in Sinhala. Integrating fine-tuned You Only Look Once version 11 (YOLOv11) for real-time abacus recognition and pre-trained Spatial Transformer Network (STN) with Residual Network 50 (ResNet50) for analog clock analysis, the system bridges physical manipulatives with digital feedback. Optimized for low-cost hardware via post-training quantization, it reduces model sizes by up to 56.6% and inference times by 45.9%, enabling deployment on standard PCs. The methodology employs YOLOv11 to detect abacus beads (mean Average Precision, mAP50-95: 0.898) and STN + ResNet50 to correct clock perspectives (95.2% accuracy on SynClock), delivering immediate Sinhala feedback through a culturally adapted interface. The system was evaluated on an Intel Core i5 PC with 8GB RAM and achieved sub-second inference (YOLOv11: 170.1 ms, STN + ResNet50: 32 ms post-quantization) while retaining accuracy. A pilot study with 20 students showed a 25% improvement in arithmetic scores and 30% in time-telling accuracy, with 85% reporting higher engagement. These findings demonstrate the system's efficacy in boosting numeracy and interaction in resource-scarce settings, aligning with the Visual, Aural, Read/Write, Kinesthetic (VARK) framework. Despite challenges like hardware constraints and real-world variability, this scalable, offline-capable solution offers a novel approach to educational technology, with potential for broader deployment in Sri Lanka and similar contexts, addressing gaps in artificial intelligence-driven kinesthetic learning tools. ©
  • Thumbnail Image
    ItemEmbargo
    CocoSense: AI-Powered Drone-Based System for Comprehensive Coconut Tree Health Monitoring and Yield Prediction
    (Institute of Electrical and Electronics Engineers Inc., 2026-05-22) Subasinghe, M; Panditharathne, R; Pasanjith, R; Nadun, T; Samarakoon, U; Tissera, W
    Coconut cultivation is vital to Sri Lanka's agricultural economy, yet farmers face significant challenges in early pest detection, disease diagnosis, and yield prediction. This research presents CocoSense, an AI-powered mobile application integrated with IoT technology for automated coconut tree health monitoring using drone-captured imagery. The system comprises four modules: (1) pest detection using EfficientNetB0 (91.44% accuracy) and MobileNetV2 (96.08% accuracy) with a trilingual AI chatbot for treatment recommendations; (2) disease detection for leaf rot, leaf spot, and leaf dieback classification (98.69% accuracy); (3) health assessment for leaf (93.70%) and branch health (99.63%); and (4) coconut yield estimation (87.86% accuracy) using YOLOv8 with dual-view acquisition strategy. Additionally, a coconut bunch detection (88.96% accuracy) module is developed to support yield estimation by identifying fruit clusters within tree canopies. The system integrates IoT-based GPS tracking with Google Maps API for real-time plantation visualization. Experimental results demonstrate that CocoSense provides a robust, accessible solution for intelligent coconut plantation management in Sri Lanka.
  • Thumbnail Image
    ItemEmbargo
    AI-Driven Vehicle Valuation and Market Trend Analysis for Sri Lanka's Automotive Sector
    (Institute of Electrical and Electronics Engineers Inc., 2025) De Silva K.P.N.T.; Shehan H.A.; Jayawardhane A.S; Premarathne A.P.S.; Krishara, J; Wijendra, D.R
    The automotive sector in Sri Lanka faces challenges in vehicle valuation accuracy and market trend analysis due to fluctuating prices, varying vehicle conditions, and environmental concerns. This paper presents an AI-driven vehicle valuation system integrating machine learning models for automated vehicle identification, damage detection, market trend analysis, and environmental sustainability assessments. Using deep learning techniques such as Convolutional Neural Networks (CNNs) and time-series models like Long Short-Term Memory (LSTM), the system delivers accurate valuation and market trend insights. Experimental results demonstrate 9 2% accuracy in damage classification and a mean absolute error (MAE) of 5.3% in repair cost estimation, supporting informed decision-making. This research bridges gaps in valuation transparency and sustainability in emerging automotive markets.
  • Thumbnail Image
    ItemEmbargo
    Intelligent Detection of Corporate Targeted Phishing Emails - A Hybrid Approach Combining Deep Learning Models with Domain Anomaly Detection
    (Institute of Electrical and Electronics Engineers Inc., 2025) Seethawaka, R; Chathurya N.E.G; Chandrasiri D.K.W.G.G.T; Kavithma K.A.S; Fernando, H; Wijesooriya, A
    This paper introduces a system designed to detect corporate-targeted phishing emails by combining two key strategies: advanced email content analysis and domain similarity analysis. The system first examines the text of emails using a hybrid deep learning model that merges modern language understanding techniques with sequential pattern recognition, achieving high accuracy in identifying phishing intent. Two models were tested - a standalone Bi-LSTM sequential model and a hybrid version(BERT - Bi-LSTM) with the hybrid model proving superior, scoring an F1 score of 0.97 compared to 0.93 for the standalone model. Second, the system verifies sender domains to detect spoofing attempts, such as subtle typos, homograph attacks or TLD/subdomain spoofing. This domain check reduces reliance on text analysis alone, helping analysts prioritize threats more effectively. Tested against a mix of legitimate and malicious domains, the domain module achieved near-perfect accuracy, minimizing false alarms. By integrating these approaches, the system addresses a critical gap in existing methods, which often focus on only one aspect of phishing (e.g., email content or URL features). This dual strategy ensures a more comprehensive defense, particularly against sophisticated attacks that use convincing language paired with fake domains. The combined model not only improves detection accuracy but also supports security teams by providing clear, actionable insights, making it practical for real-world corporate environments.
  • Thumbnail Image
    ItemEmbargo
    A Dual-Branch CNN and Metadata Analysis Approach for Robust Image Tampering Detection
    (Institute of Electrical and Electronics Engineers Inc., 2025) Zakey, A; Bawantha, D; Shehara, D; Hasara, N; Abeywardena, K.Y; Fernando, H
    Image tampering has become a widespread issue due to the availability of advanced tools such as Photoshop, GIMP, and AI-powered technologies like Generative Adversarial Networks (GANs). These advancements have made it easier to create deceptive images, undermining their reliability and fueling misinformation. To address this growing problem, we propose a hybrid approach for image forgery detection, combining deep learning with traditional forensic techniques. Our study integrates a dual-branch Convolutional Neural Network (CNN) with handcrafted features derived from Error Level Analysis (ELA), noise residuals from the Spatial Rich Model, and metadata analysis to enhance detection capabilities. Metadata analysis plays a crucial role in identifying inconsistencies in image properties such as timestamps, geotags, and camera details, which often accompany tampered images. The CASIA dataset, a publicly available benchmark for tampered images, was used to train and evaluate the proposed model. After 30 epochs of training, the hybrid method achieved an accuracy of 95%, demonstrating its effectiveness in distinguishing between authentic and tampered images. This research highlights the advantages of combining deep learning models with traditional feature extraction methods and metadata analysis, offering a robust solution for detecting manipulated images. Our findings contribute to advancing image forensics by improving detection accuracy, even in cases involving sophisticated tampering methods driven by AI.