Research and Publications
Published - 4 ( First Author -3)
Ongoing research - 2
Ongoing collaboration with students - 1
Published - 4 ( First Author -3)
Ongoing research - 2
Ongoing collaboration with students - 1
Ongoing Research Endeavors
As a member of the Data Analysis and Deep Learning Laboratory, my current research focuses on advancing the Genomic Foundation Model through novel tokenization strategies and scalable pre-training under the supervision of Dr. Aminul Islam, Associate Professor of Computer Science, University of Louisiana at Lafayette.
I am also working on Prototype-Guided Learning for Knee Osteoarthritis Classification and Synthetic Data Generation.
Published Works
A first-of-its-kind system to segment words from Bangla text images using novel Image Processing and Computer Vision algorithms, addressing challenges such as variations in page types, shadow interference, backgrounds, ink colors, and handwriting styles and sizes.
New Terminology such as Minimum Area Threshold, Writing Density , and Average Contour Size was introduced to generalize a text image's content.
Tested on two conventional datasets and one new dataset, the system achieved an impressive F1-score of 91.20%.
Developed an Open Vocabulary word recognition system using an ensemble of two Deep Learning models, complemented by unique Computer Vision techniques. The system includes a modified Non-Maximum Suppression algorithm that enables recognition of words, even if they do not exist in the Bangla dictionary, making it adaptable to other languages.
Introduced new terminology such as, overlap value, to address unresolved challenges.
Tested on two conventional datasets and one new dataset, the ensemble method outperformed traditional approaches, achieving an impressive F1-score of 92.61%.
Developed a novel Dissimilarity Matrix for Bangla OCR error correction, acting as an intermediary layer between Levenshtein Distance and contextual analysis. The method significantly improves OCR accuracy by handling visual similarity errors between Bangla characters from the machine’s perspective.
With a strengthened dictionary and enhanced Edit Distance Matrix, the system achieved an impressive F1-score of 93.04%.
Developed two novel datasets through the efforts of a team of over ten people working for six months. The datasets are published in a respected Elsevier repository, and a detailed manuscript is being prepared for journal submission.
Collaboration with Students- Currently under development.
At present, I am supervising a group of eighth-semester students from the University of Information Technology & Sciences engaged in Capstone Projects. We are developing “BDTourAI: A Fine-Tuned LLM for Bangladesh Tourism,” which provides detailed travel plans for tourists exploring Bangladesh. A dataset has been created with 50+ highly visited locations, incorporating features like weather, distance from the capital, and cost of food.