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
Continuous sign language interpretation to text using deep learning models
(Institute of Electrical and Electronics Engineers Inc., 2022-01-01) Rahman, Afridi Ibn; Akhand, Zebel-E-Noor; Nahian Khan, Tasin Al; Sarda, Anirudh; Bhuiyan, Subhi; Rakib, Mma; Ahmed Fahim, Zubayer; Kundu, Indronil
The COVID-19 pandemic has obligated people to adopt the virtual lifestyle. Currently, the use of videoconferencing to conduct business meetings is prevalent owing to the numerous benefits it presents. However, a large number of people with speech impediment find themselves handicapped to the new normal as they cannot communicate their ideas effectively, especially in fast paced meetings. Therefore, this paper aims to introduce an enriched dataset using an action recognition method with the most common phrases translated into American Sign Language (ASL) that are routinely used in professional meetings. It further proposes a sign language detecting and classifying model employing deep learning architectures, namely, CNN and LSTM. The performances of these models are analysed by employing different performance metrics like accuracy, recall, F1- Score and Precision. CNN and LSTM models yield an accuracy of 93.75% and 96.54% respectively, after being trained with the dataset introduced in this study. Therefore, the incorporation of the LSTM model into different cloud services, virtual private networks and softwares will allow people with speech impairment to use sign language, which will automatically be translated into captions using moving camera circumstances in real time. This will in turn equip other people with the tool to understand and grasp the message that is being conveyed and easily discuss and effectuate the ideas.
Multimodal emotion recognition using heterogeneous ensemble techniques
(Institute of Electrical and Electronics Engineers Inc., 2022-01-01) Esfar-E-Alam, A.M.; Hossain, Mehran; Gomes, Maria; Islam, Rafidul; Raihana, Ramisha; Department of Computer Science and Engineering
Emotion recognition and sentiment analysis serve several purposes, from analyzing human behavior under specific conditions to the enhancement of customer experience for various services. In this paper, a multimodal approach is used to identify 4 classes of emotions by combining both speech and text features to improve classification accuracy. The methodology involves the implementation of six models for both audio and text domains combined using four different heterogeneous ensemble techniques - hard voting, soft voting, blending and stacking. The effects of each ensemble method on the accuracy for the multimodal classification task are also investigated. The results of this study show that the usage of ensemble learning to combine modalities greatly improves classification, with stacking being the best-performing ensemble technique for the selected collection of models. The proposed model outperforms several existing methods for 4-class emotion detection on the IEMOCAP dataset, obtaining a weighted accuracy of 81.2%.
Non-Fungible Token (NFT): Analyzing marketplaces and non-user perspectives
(Institute of Electrical and Electronics Engineers Inc., 2022-01-01) Saifullah S.I.; Islam S.; Ferdous, Md. Sadek; Chowdhury F.; Department of Computer Science and Engineering
Non-Fungible Tokens (NFTs) have emerged as a disrupting revolutionizing technology, particularly for the digital art and collectibles industries. Understandably, they have received a huge attention from the digital as well as traditional artists. Powered by smart-contracts deployed in different blockchains, NFTs have provided more control and power to the content creators than ever before and gave birth to the core concept of verifiable digital ownership. These NFTs are traded in different online marketplaces. In order to ensure a wide-scale adoption of NFTs, it is important to identify potential hurdles that might hamper their wide scale adoption. Towards this aim, we present the findings of two studies in this paper. Firstly, we present an evaluation framework for different NFT marketplaces, which clearly illustrates the difference between their offered features. Secondly, we carry out a semi-structured interview to identify common drawbacks that are holding people back from NFT adoption. Our results reveal that the whole technology backing NFTs need to be revised for wider adoption. We sum up our findings from different point of views where the barriers exist but viable improvements are suggested for better and far-reaching promotion of NFTs.
Classification of damaged vegetation areas using convolutional neural network over unlabelled sentinel-2 images
(Institute of Electrical and Electronics Engineers Inc., 2021-01-01) Haque, Samiha; Rahman, Nazibur; Mostakim, Moin; Department of Computer Science and Engineering
Several researches have been made in recent years which use Convolutional Neural Networks (CNN) to classify multispectral images. However, a scarcity of labelled satellite images makes it difficult to classify damaged forest areas using pre-trained CNNs. If any new destruction hits the forests, then there will be a lack of labelled satellite data and manually labeling the satellite images will be inefficient and will ultimately fail the purpose of rapid reforestation. Hence, this research aims to label the unlabelled Sentinel-2 satellite images by using a pre-trained ResNet-50 to obtain flattened feature maps from difference images of disaster struck vegetation areas followed by applying K-Means clustering algorithm on them. Thus, a complete labelled dataset is produced which is used to train two CNN models to classify the areas into undamaged and damaged classes for Model-1 and undamaged, mildly damaged and severely damaged classes for Model-2. Model-1 achieved 89.77% test accuracy and Model-2 achieved 85.69% test accuracy. Model-1 recorded 89.77% in micro f1-score, 76.99% for kappa score and 0.10 for overall error and Model-2 had 85.69%, 72.87% and 0.14 values in micro f1-score, kappa score and overall error respectively.
The eigenvalue distribution of Hankel matrix: A tool for spectral estimation from noisy data
(Institute of Electrical and Electronics Engineers Inc., 2021-01-01) Islam A.; Hasan, Md. Rakibul; Hossain M.Z.; Hasan M.M.; Department of Electrical and Electronic Engineering
One of the key challenges of digital signal processing is to estimate sinusoidal components of an unknown signal. Researchers and engineers have been adopting various methods to analyze noisy signals and extract essential features of a given signal. Singular spectrum analysis (SSA) has been a popular and effective tool for extracting sinusoidal components of an unknown noisy signal. The process of singular spectrum analysis includes embedding time series into a Hankel matrix. The eigenvalue distribution of the Hankel matrix exhibits significant properties that can be used to estimate an unknown signal's rhythmic components and frequency response. This paper proposes a method that utilizes the Hankel matrix's eigenvalue distribution to estimate sinusoidal components from the frequency spectrum of a noisy signal. Firstly, an autoregressive (AR) model has been utilized for simulating time series employed to observe eigenvalue distributions and frequency spectrum. Nevertheless, the approach has been tested on real-life speech data to prove the applicability of the proposed mechanism on spectral estimation. Overall, results on both simulated and real data confirm the acceptability of the proposed method. This study suggests that eigenvalue distribution can be a helpful tool for estimating the frequency response of an unknown time series. Since the autoregressive model can be used to model various real-life data analyses, this study on eigenvalue distribution and frequency spectrum can be utilized in those real-life data. This approach will help estimate frequency response and identify rhythmic components of an unknown time series based on eigenvalue distribution.