Lecturer, Department of Software Engineering, Sabaragamuwa University of Sri Lanka · Research Intern, University of Southern Queensland
My research interests lie in computer vision, hybrid quantum-classical machine learning, and explainable AI, with applications in medical imaging, precision agriculture, and 3D computer vision. My current work focuses on generative medical image evaluation, image segmentation, object detection, visual SLAM, and 3D reconstruction.
I hold a B.Sc.Eng. (Hons) in Computer Engineering from the University of Peradeniya.
Scholarship
Papers where I am a primary or significant contributor. Full record on Google Scholar and ResearchGate.
Synthesises state-of-the-art deep learning techniques, benchmarking trends, and open challenges for citrus leaf disease recognition in real-world deployment.
Semi-supervised framework using pseudo-labelling to classify tomato leaf diseases under label scarcity, sharply reducing dependence on annotated data.
Robust Vision Transformer framework for early tomato leaf disease detection using attention-based feature extraction.
Systematic study of amplitude vs. angle encoding and qubit scaling in hybrid quantum-classical CNNs for dermatological skin lesion classification.
Comprehensive survey categorising semi-supervised methods that reduce dependence on large expert-annotated medical datasets.
_page-0001.jpg)
Ultra-lightweight federated learning with attention for privacy-preserving disease recognition on edge devices.
Ultra-tiny Vision Transformer with an affix attention mechanism for accurate image classification on mobile and resource-constrained devices.
Large-scale, high-resolution brinjal leaf disease dataset captured in real field conditions for robust model benchmarking.
Survey of deep learning for physiological signals, focusing on multimodal fusion strategies and explainability in healthcare AI.
Benchmarks ten XAI techniques for breast cancer imaging across CNN, ViT, and hybrid architectures.
Self-supervised ViT with built-in explainability for disease detection without labelled training data.
Explainable few-shot ViT with metric-based learning for breast cancer detection from limited labelled mammography data.
Ultra-lightweight hierarchical dilated convolution network for real-time disease recognition on edge devices.
Unified Zero Trust security framework for Kubernetes integrating SPIRE, Cilium, Falco, Kyverno, and Vault.
Cross-layer event correlation and automated response validation for Zero Trust Kubernetes deployments.
One of the largest controlled XAI benchmarks for modern vision architectures — 13 attribution methods across CNNs and five ViT families, evaluated on faithfulness, localization, robustness, complexity, and efficiency.
HiLRP: a unified explainability framework with conservation-valid relevance propagation for arbitrary ViTs, decomposed into four attention primitives.
Evaluation metric built on a frozen radiology-pretrained vision transformer, combining CKA-based layer selection, entropy weighting, and multi-scale MMD.
Hybrid quantum-classical architecture mirroring autonomic nervous-system subsystems for physiological signal recognition.
Learned spatial superpixel attention conditioning a trainable quantum encoding — up to 93.9% accuracy across seven benchmarks (incl. PathMNIST, PneumoniaMNIST) with as few as 60K trainable parameters, a ~100× reduction versus ResNet-18.
Focus areas
Object detection, semantic segmentation, SLAM, Vision Transformers, self- and semi-supervised learning.
Hybrid quantum-classical architectures, variational quantum circuits, encoding strategies and qubit scaling.
Attribution methods (Grad-CAM, LIME, SHAP), faithfulness evaluation, and model transparency for vision models.
TinyML, ultra-lightweight architectures, and privacy-preserving distributed training on constrained devices.
Precision agriculture (crop disease recognition) and medical imaging (histopathology, dermatology, mammography).
Background
Instruction
Mentoring
Final-year research projects I supervise in the Department of Software Engineering, Sabaragamuwa University of Sri Lanka.
N. A. Rajapaksha · Final-year research project
A framework that scrapes 10,000+ articles from Sinhala news websites, classifies events with Sinhala-BERT, groups related coverage using UMAP and HDBSCAN with hierarchical event clustering, and generates event-level extractive summaries with TextRank — evaluated on classification metrics, clustering quality, and human judgment of summary quality.
L. D. Weerakoon · Final-year research project
Benchmarks a custom Mamba state-space classifier against SinhalaBERTo, IndicBERT, and mBERT for sarcasm detection on ~3,000 annotated Sinhala YouTube comments. SinhalaBERTo leads thanks to language-specific pretraining, while the Mamba model outperforms mBERT — evidence that state space models are a feasible alternative for low-resource Sinhala NLP.
T. P. H. Nethmin · Final-year research project
A cloud-native pricing framework forecasting parking demand from historical transactions, weather, and event data. Seven models — gradient-boosting ensembles, LSTM variants, CNN-LSTM, and a Transformer — were compared on forecasting accuracy, revenue, price stability, and operational fairness; the Transformer performed best overall, and the deployed Azure solution achieved a 100% API success rate.
T. D. Rathnasuriya · Final-year research project
Tackles temporal inconsistency in audio-visual segmentation — flickering masks between video frames — using AVSegFormer on the AVSBench single- and multi-source benchmarks. Introduces consistency-aware evaluation (Mask Consistency Score, Warp IoU) alongside mIoU and F-score, and compares four remedies; recurrent query propagation improves both accuracy and temporal stability, especially on the harder multi-source videos.
S. A. G. N. Jayasekara · Final-year research project
Quantifies the privacy-utility trade-off in public surveillance by comparing anonymization techniques — Gaussian blurring, pixelation, masking, and a novel Pose-Conditioned Silhouette approach — across convolutional (YOLOv8) and transformer-based (RT-DETR) detectors on the WiderPerson dataset, measuring privacy via SSIM and detection accuracy via mAP50.
Open source & my leisure time builds
Selected research code and side projects. More on GitHub.
Quantum circuit simulator with real-time Bloch sphere visualization, VQE, and Grover's Search.
Unified codebase implementing state-of-the-art semi-supervised image recognition algorithms.
Her2 multi-class breast cancer cell analysis with explainable AI attribution mapping.
Real-time emotion detection architectures optimized for low-power edge microcontrollers.
Multi-modal stress and physiological state detection from WESAD wearable biosensor signals.
Lightweight MobileNetV2 adaptation for fast 1D physiological signal pattern recognition.
Skin cancer dermoscopy dataset analysis with computer vision classification pipelines.
Dashboard to control, process, and analyze 3D photorealistic scans via Luma AI APIs.
Computer vision framework translating hand gestures and sign language in real time.
Browser-based presentation app rendering AI-generated lecture structures and animations.
Interactive non-communicable disease data dashboards using Node.js and D3.js.
Verification framework comparing PDF structural layouts and content changes in pipelines.
Automated pipeline to generate, edit, and assemble dynamic video content with audio.
Web application classifying user reviews and feedback using NLP pipelines.
Intelligent router analyzing and categorising student feedback transcripts.
Full-stack student connection platform with posts, feeds, and event scheduling.
Virtual line-following robot simulation over customizable image-based tracks.
