Md Bikasuzzaman

Md Bikasuzzaman, Alif Arman, Ishtiaq Hossen, Arindam Kishor Biswas, Alauddin Sabari

IEEE Conference on QPAIN, 2025Published

Guava (Psidium guajava) produces a colorful fruit full of fiber, vitamins, and antioxidants. This study aims to categorize several guava diseases so that producers can take preventive measures. We apply both deep learning and conventional machine learning techniques for early detection. Two distinct datasets are used from the Mendeley Data website. InceptionV3 is used for feature extraction, PCA for dimension reduction, and the K-Nearest Neighbors (KNN) method is used for classification, which produces a precision rate of 96.83%. With the help of VGG19 and KNN, we extract features with an average accuracy of 69% for the second dataset.

TinyViT-HAR: Human Activity Recognition via WiFi CSI Signals by Lightweight Vision Transformers

Md Shafiqul Islam, Mithila Farjana, Mustakim Musully Pias, Md Bikasuzzaman, Md Junaid Hossain, S.M. Abdur Rahim

IEEE RAAICON, 2025Accepted

Human Activity Recognition (HAR) has become essential in various intelligent applications, including smart homes, healthcare, and security. Traditional HAR methods often rely on wearable sensors or cameras, which pose challenges related to privacy, user discomfort, and limited applicability in complex environments. To overcome these limitations, this paper proposes TinyViT-HAR, a novel lightweight Vision Transformer-based framework that utilizes WiFi Channel State Information (CSI) signals for activity recognition. By treating CSI data as image-like inputs, the model effectively captures rich spatiotemporal patterns using a multi-head attention mechanism. To enhance efficiency, we introduce structured pruning, significantly reducing model complexity without sacrificing accuracy. Experimental results on public WiFi CSI datasets demonstrate that TinyViT-HAR achieves up to 99.28% accuracy, outperforming existing models while maintaining a compact size suitable for edge deployment. Even under high pruning ratios, the model preserves robustness, achieving 98.92% accuracy with just 57K parameters. The proposed framework offers an effective, non-intrusive, and resource-efficient solution for real-time HAR in practical environments.

Semantic Dissection of Bengali Feedback: Benchmarking Hybrid Deep Neural Networks for Multilabel E-Commerce Classification

M. Bikasuzzaman, A. Arman, N. Nowshin

IEEE Conference on QPAIN, 2026Under Review

A Transformer-Based Approach for Summarizing Employee Logs

Md Bikasuzzaman, Alif Arman, Ishtiaq Hossen, Arindam Kishor Biswas, Alauddin Sabari

Under Review, 2025Under Review

Efficient summarization of employee daily logs is crucial for organizational productivity. This study utilizes the Google Pegasus model to enhance summarization, addressing traditional challenges with complex document structures. A dataset of daily work logs from a software development firm was augmented with advanced Generative AI techniques. The Pegasus model was fine-tuned with an input token length of 1024 and an output token length of 128. Evaluated using the ROUGE metric, the model achieved a ROUGE-1 score of 0.613, ROUGE-2 score of 0.373, ROUGE-L score of 0.557, and ROUGE-Lsum score of 0.556, demonstrating strong performance in generating coherent and contextually relevant summaries.