Research Experience

We designed an innovative dehydration detection system that leverages a multi-modal Siamese network to analyze hand vein images captured by smartphone IR cameras. The system includes two distinct ROI extraction methods: a unique geometry-based algorithm and the YoloV8 segmentation model. Additionally, we implemented an image processing step for vein segmentation within the identified ROI and developed a novel process for selecting the optimal frame from hand-vein videos, ensuring accurate and reliable results.

We developed a novel technique to detect advanced web bots, specifically those driven by OpenWPM, using behavioral analysis. By collecting client-side data such as mouse movements, keystrokes, and scrolling patterns, we clustered this data using k-means and trained three machine learning models—Random Forest, Linear Regression, and Support Vector Machine—on each cluster. These models are then used to classify human and bot behaviors in real-time. Our method achieves a 99.1% accuracy rate while minimizing false negatives, making it effective for identifying sophisticated bots that mimic human browsing behavior

For the Bangla Sign Language Production project, we created a new sentence-level dataset. We are now developing a progressive transformer model to improve the accuracy and fluency of Bangla sign language by converting skeleton key points into meaningful sequences.