Publications

Non-invasive microwave glucose sensing produces subtle, nonlinear, and noise-sensitive changes in S-parameters, making glucose concentrations difficult to distinguish reliably using raw microwave measurements alone.

Combined microwave S-parameter analysis with machine-learning and signal-processing techniques. Statistical and interaction-based feature engineering was applied to magnitude and phase measurements, followed by DWT-based Sensor Pattern Noise extraction and comparison of ensemble classifiers including Random Forest, Gradient Boosting, XGBoost, LightGBM, and Extra Trees across simulated and measured datasets.

Feature engineering and wavelet-based Sensor Pattern Noise substantially improved glucose-level classification compared with raw S-parameters. The proposed SPN stage achieved up to 80.6% accuracy on the simulated dataset, while Extra Trees reached 73.27% on measured data, demonstrating improved discrimination under realistic measurement variability.

Demonstrated how signal processing and machine learning can extract glucose-dependent patterns that are difficult to identify directly from raw microwave measurements, providing a data-driven path toward more robust non-invasive glucose sensing while identifying remaining real-world validation and deployment challenges.

Multimodal biosensing can provide richer physiological information than single-sensor systems, but practical deployment is limited by heterogeneous signals, synchronization and fusion complexity, power constraints, environmental interference, and computational overhead.

Reviewed wearable biosensing modalities and examined how signal processing, multimodal data fusion, and AI algorithms can connect physical sensors with on-device and system-level intelligence. The review covers lightweight edge models, deep learning, reinforcement learning, and multi-stage fusion strategies for intelligent healthcare.

The review establishes a structured sensor-computing perspective spanning on-sensor processing, on-device intelligence, and system-level fusion, and identifies practical AI pipelines for preprocessing, feature extraction, multimodal fusion, prediction, and validation.

Provides a practical framework for moving multimodal biosensors from isolated sensing devices toward intelligent sensor-computing systems, while highlighting the need to address real-world domain shift, explainability, privacy, and deployment constraints before clinical translation.

Low-resolution and degraded license-plate images reduce recognition reliability in unconstrained surveillance environments.

Built an SRGAN-based super-resolution pipeline with modified network architecture, Swish activation, and TV regularization, followed by YOLOv5-based plate-character recognition.

The proposed models improved reconstructed-image quality and achieved up to 95.4% recognition accuracy, while outperforming comparison methods in PSNR and SSIM.

Demonstrated that image super-resolution can improve downstream ALPR performance rather than treating image enhancement and recognition as separate problems.

Automatic license-plate recognition in unconstrained video requires both high detection accuracy and sufficiently low inference latency for practical real-time and edge-oriented deployment. Traditional handcrafted features also have limited generalization across variations in lighting, image quality, viewpoint, and scene conditions.

Built and evaluated a deep-learning license-plate detection workflow using SSD, Faster R-CNN, and YOLO architectures. The study included custom data collection and bounding-box annotation, model-specific training and hyperparameter configuration, and evaluation on video sequences to compare detection accuracy and inference response time.

YOLO achieved the strongest tested balance of accuracy and response time, reaching 97% accuracy with approximately 2 s prediction time in the reported setup. Faster R-CNN achieved 91% at approximately 2.5 s, while SSD achieved 44% at approximately 3.6 s.

Demonstrated the practical trade-off between detection accuracy and response time when selecting convolutional object-detection architectures for video-based ALPR, providing an experimental basis for choosing models for resource-constrained and Edge AI-oriented systems.

Mesh-based video coding for ultra-high-definition content requires motion estimation that preserves visual quality without excessive computational cost. Different three-step-search strategies can produce substantially different trade-offs between reconstruction quality and matching time when used with HASM.

Compared Ordinal, Diamond, Grid, and Leap Three Step Search motion-estimation methods within a HASM-based hierarchical mesh workflow. The algorithms were evaluated under common experimental conditions on four 4K video sequences, using PSNR to measure reconstructed visual quality and average matching time per frame to characterize computational performance.

Diamond and Grid TSS provided the strongest visual-quality results while maintaining matching times close to the faster Ordinal TSS. Ordinal TSS offered a practical quality-speed compromise, whereas Leap TSS generally produced the weakest combined quality and timing behavior in the tested configurations.

Shows that motion-estimation strategy materially affects both reconstructed quality and computational cost in HASM-based UHD video processing, providing practical guidance for selecting search algorithms according to whether visual quality or computational efficiency is the dominant constraint.

Hierarchical Adaptive Structured Mesh (HASM) had been proposed for scalable mesh-based motion representation, but its behavior on ultra-large video formats such as 1080p, 4K, and 8K had not been systematically evaluated, particularly the relationship between reconstructed visual quality and mesh granularity.

Evaluated HASM-based motion tracking using standard high-resolution video sequences and multiple fixed and hierarchical mesh configurations. Motion vectors were estimated from successive frames, predicted frames were reconstructed from the reference frame and motion vectors, and visual quality was assessed using PSNR across different initial and minimum mesh sizes.

Evaluation across the Bosphorus, Jockey, and HoneyBee sequences showed consistent reconstruction quality across the tested HASM configurations. Increasing the level of mesh splitting provided finer motion representation but also increased bitrate requirements, exposing a practical trade-off between visual fidelity, mesh detail, and coding overhead.

Provides an empirical basis for configuring HASM for large-format video by showing how mesh resolution affects reconstruction quality and bitrate, supporting more informed selection of hierarchical mesh parameters for high-resolution video coding.

Static resource and power-allocation strategies cannot adapt efficiently to the strong temporal variation in mobile-network traffic, which can lead to over-provisioned transmission resources and unnecessary energy consumption.

Derived normalized weekday and weekend mobile-traffic signatures from CDR activity, clustered sites with similar temporal behavior using K-means, and used the resulting traffic groups to support dynamic resource allocation and traffic-dependent power modeling. Power profiles were estimated using curve-fitting and regression-based models.

The traffic-signature approach identified distinct site-behavior clusters and enabled dynamic transmission-resource allocation, reducing required link capacity by 19.17% compared with traditional peak-based planning. The cluster-based power models also achieved strong fitting performance, with reported power-estimation accuracy of approximately 96%.

Demonstrated that learning recurring traffic behavior can support more adaptive network planning, allowing transmission resources and power consumption to follow actual demand rather than fixed peak assumptions.