BRAC University Institutional Repository
Preserving Knowledge, Advancing Research, Sharing Scholarship
A digital platform for collecting, preserving, and sharing BRAC University’s scholarly, academic, and institutional outputs.
Contribute to the Repository
Faculty members and students are invited to submit their research publications, theses, dissertations, and scholarly works to increase visibility, access, and long-term preservation.
Recent Submissions
BDCoins: A comprehensive dataset for Bangladeshi coin detection using YOLOv11
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Iqbal, Khondoker Nazia; Taj, Towshik Anam; Mahee, Md Nafiz Ishtiaque; Fahim, Mohammad; Zereen A.N.; Department of Computer Science and Engineering
Currency detection is a complex task due to the diverse patterns and rich features found in different currencies. Identifying coins presents unique challenges, as their appearance can vary with orientation and environmental conditions. Recent approaches in the field shift from manual feature engineering to automated systems using deep learning, which demonstrate superior accuracy and robustness. Object detection models like YOLO have become popular for coin recognition due to their speed and accuracy, with research works applying various versions to identify specific national currencies. Despite these advances, research on Bangladeshi currency detection, particularly for coins, remains very limited. A significant research gap exists because there is no large, publicly available dataset that includes the newly designed 1, 2, and 5 Taka coins, their variations, and images of both their front and back sides. This paper addresses this gap by introducing BDCoins, a custom benchmark dataset containing 11,133 annotated images of Bangladeshi coins. The dataset encompasses all old and new variations of the 1,2, and 5 Taka denominations, with images captured under diverse conditions to reflect real-world scenarios. A YOLOv11 model is trained and validated on this dataset for detection and classification. The model achieves an F1 score of 0.983 and demonstrates 0.982 accuracy in testing, providing a foundational tool for automated Bangladeshi currency recognition systems.
Explainable AI framework for SNR prediction and adaptive beamforming in mmWave 5G networks
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Singha S.; Arnob I.A.; Chowdhury A.I.; Reza, Rifat Bin; Bhadra S.; Chowdhury A.; BSRM School of Engineering
Extremely fast development of the fifth-generation (5G) mobile networks, and in millimeter-wave (mmWave) issues, adaptive beamforming methods must be highly efficient to ensure stable signal-to-noise ratio (SNR) in various propagation conditions. The current paper describes an explainable artificial intelligence (XAI)-based system of SNR prediction that combines the domain knowledge of the antenna theory with machine learning models. A description database was designed with the implementation of domain-specific features, which included the essential parameters like array structure, propagation conditions, and side-lobe, and the use of RF chains. The wide scope of exploratory data analysis and statistical confirmation, such as ANOVA and ANCOVA, demonstrated that the behavior of SNR is very nonlinear and cannot be explained by point-specific factors. Then, LazyRegressor was used to compare the performance of multiple regression models, and feature engineering significantly improved the performance of the models by changing the values of the R2 to 1.00. Interpretability of the models was further improved with the help of SHAP analysis which allowed seeing the contribution of the parameters clearly. The suggested framework does not only go further to promote the accuracy of prediction but also forms an interpretable decision-supporting adaptive beamforming in mmWave 5 G systems. These results reveal the significance of domain knowledge and explainability in the development of next-generation wireless communication solutions.
Generative AI meets responsible AI and affective computing
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Shomrat, Kamran Hassan; Islam, Gazi Arman; Eva, Atkea Fauzia; Islam, Md. Saiful; Subarna, Jamilatun; Tahsin, Anika; Alam, Md. Golam Rabiul; Department of Computer Science and Engineering
Artificial Intelligence (AI) systems are increasingly expected to interpret and respond to human emotions in real time, yet most existing solutions remain limited to unimodal or ethically unregulated approaches. In this research, we propose a unified, multimodal emotion recognition (MER) framework that integrates Generative AI (GenAI) and Responsible AI (RAI) with Affective Computing to enhance human-computer emotional interactions. We explore real-time emotion recognition and response by analyzing image, speech, and text, while also incorporating user gestures and behavioral signals. We construct a custom dataset of 1000 annotated video samples spanning seven emotion classes. We built our proposed approach using transformer-based models, including DistilRoBERTa, fine-tuned Wav2Vec2, and DeepFace to process text for sentiment classification, audio for emotion detection, and facial image modalities for facial expression analysis. A majority-vote fusion strategy combines the outputs of these models to identify dominant emotional states and gradually learns from user feedback, adapting to behave more human-like and empathetic. Additionally, our system also incorporates RAI layers to ensure ethical safeguards through bias and threat detection. Finally, a FLAN-T5-based generative module produces natural language summaries that reflect both the emotional content and ethical assessments. Our proposed method achieves an overall accuracy of 81%, with F1-scores of 0.94 and 0.90 for anger and disgust, respectively. Our approach suggests the potential for significant improvements beyond unimodal baselines, enabling ethically aware and emotionally intelligent applications such as virtual assistants, mental health support systems, and emotion-adaptive learning platforms.
SBIoT-ATM: Secure blockchain-IoT framework for next-generation automated teller machine security
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Nafis, Farhan Ahmad; Fahomida, Maiesha; Sakib Alvi, Md. Saadman; Salim Rafid, Sk Tahmed; Department of Computer Science and Engineering; Department of Electrical and Electronic Engineering
Modern infrastructure requires state-of-the-art security systems; as system implementation advances, the threats of cyberattacks and intrusions also evolve. Automated Teller Machines (ATMs) are a kind of financial structure that provides a crucial service to customers. To address security concerns using cutting-edge technology today, this paper explores the possibility of incorporating monitoring nodes and IoT with blockchain. We systematically explore the effect of combining these feature sets and propose our system. We evaluate our system using numerous testing procedures, including encryption and decryption capabilities, node network optimization, and a blockchain-based security system. We ran tests on our proposed system and achieved an average throughput of 0.93264 Mbps, with an average delay of 22.64 ms and a PDR of 97.9%. Our FOG security scheme achieved encryption and decryption times of 0.2134 ms and 0.2003 ms, respectively, while maintaining CPU usage below 12%. Our lightweight blockchain used 96.33 MB and 109.33 MB, with CPU usage of 11.5% and 23.1%, during the computation of the second block and subsequent blocks after the genesis block.
An advanced machine vision technique for quality control of fabric in the textile industry
(Institute of Electrical and Electronics Engineers Inc., 2025-01-01) Chowdhury, Rifat Arman; Bhattacharjee, Souharda; Sarker, Sajib; Department of Computer Science and Engineering
The textile industry in Bangladesh, one of the leading sectors of the country's economy, requires flawless product quality to ensure their worldwide reputation. However, one of the major setbacks they face is the detection of defects in fabric. Traditionally, quality control has relied on manual inspection, which is timeconsuming, inefficient, and prone to human error; which incurs significant financial losses to the industry. Fabric defects introduced during manufacturing or processing make visual inspection necessary; however, the repetitive and monotonous nature of this task makes manual detection unreliable. This paper proposes a system to inspect faults on fabric in the production line of textile industries using machine vision techniques to ensure greater precision. We analyzed established deep learning based object detection models such as YOLOv8n, Single Shot MultiBox Detector (SSD) with VGG16 backbone, and Faster R-CNN with ResNet-50, and further developed a novel two-stage pipeline architecture. In our hybrid approach, YOLOv81 is employed for initial defect detection, followed by EfficientNet-B0 for classification. A custom dataset, containing a total of 9,075 images across four defect classes, was developed using pictures taken from textile factory environments to simulate real industrial settings. We also designed a prototype system for industrial automation to integrate our model into practical workflows. Fabric defect detection, automated using machine vision techniques is a growing research focus in the area, offering efficient, fast, and scalable solutions to quality control. While previous methods have been widely used, they often suffer from limitations in adaptability, accuracy, and real-world deployment. Our proposed model not only addresses these drawbacks, but also enables reliable inspection of fabric faults in industrial production. It is especially suitable for developing an economical and customizable system for fault detection. The pipeline achieved a precision of 0.924, recall of 0.891, and F1-score of 0.913 - demonstrating both accuracy and industrial applicability.