Thesis (Master of Science/Engineering in Computer Science and Engineering)

Browse

Now showing 1 - 20 of 101

3D model based interactive application for elementary education
(BRAC University, 2019-08) Ghosh, Apurba; Uddin, Jia
This thesis demonstrates a 3d model based interactive application which has been initially deployed for android platform and functions on the basis of augmented reality technology targeting the sector of Bangla language based elementary education. The development of this model includes an extreme level of engineering particularly focusing on polygon and vertex count while developing 3d assets in virtual environment. As this model is based on augmented reality technology, real-world triggering has a huge impact on it and the preparation of these triggers were a crucial part of this scientific endeavor. To validate the performance of this model we have tested it in five different environments based on three core matrix and figured out that this kinds of models are a great fit for students aged from 4 to 5 years old along with a staggering returning rate of 56 times on an average within 3 days. Our study also shows that the learning rate of male students are relatively faster than female students by 3.08% when they are using our proposed model as a medium of learning.
3G and 4G paging success rate based mobile network anomaly detection using supervised and unsupervised learning
(BRAC University, 2022-04) Ahasan, Md Rakibul; Alam, Md. Golam Robiul
In a mobile network, there are a lot of data that can provide network detail about network efficiency, robustness, and availability. A type of data is mobile network performance data obtained from the key performance indicators (KPI) or the key quality indicators (KQI). An integral part of mobile network monitoring is it monitor any unusual pattern in the performance data. The pattern detection or anomaly detection use case from performance data is essential for mobile operators because it detects issues in the network that are not possible to detect by the network alarms. A machine learning-based anomaly detection model is most common nowadays. This thesis demonstrates a supervised and unsupervised machine learning-based anomaly detection model. The base data set is paging success rate performance data of day-level and hourly-level granularity. Secondly, a comparative analysis is present over various anomaly detection models. Thirdly, the data used in this paper has an imbalance scenario and how the re-sampling technique can affect the outcome of the anomaly detection model. Lastly, one supervised machine learning recommends mobile network anomaly detection. However, implementing supervised machine learning over a large data set is more computational because it requires ground truth determination. On the other hand, unsupervised machine learning will cluster various data volumes without any prerequisite. If proper tuning is in place on this model, it will give an efficient anomaly detection. Another aspect of this thesis is to identify unsupervised machine learning that is best suited for mobile network anomaly detection. To do that a benchmarking approach is performed over three unsupervised machine learning, and these are K-means, DBSCAN, and HDBSCAN. The thumb rule of the benchmark follows as converting the unsupervised machine learning output into a classification problem and then measuring the model performance. The deep learning implication of anomaly detection in 4G network performance data exercise in this thesis and an autoencoder used to see how it performs in anomaly detection with moderate accuracy.
A character gram modeling approach towards Bengali Speech to text with regional dialects
(BRAC University, 2024-11) Hassan, Md. Rezuwan; Rabiul Alam, Dr. Md. Golam
The Bengali language, spoken in various regions of south-Asia and also among the Bengali diaspora, exhibits rich diversity with regional dialects or variations that reflect the cultural, geographic, and historical influences of different regional/sociocultural communities. Based on phonology and pronunciation, Bengali is said to have 5 distinct major dialectal variations, such as Eastern Bengali Dialect, Manbhumi, Rangpuri, Varendri, and Rarhi. For the dialects present in Bangladesh, even finer stratification can be done based on the used vocabulary, pronunciation, phonology, syntax, and morphology.These regional Bengali dialects are found in regions such as Bangladesh in the regions of Chittagong, Sylhet, Rangpur, Rajshahi, Noakhali, Barishal, etc possess unique phonetic, lexical, and syntactic features that set them apart from standard Bengali and also unique from each other. However, research and resources dedicated to understanding and harnessing the potential of natural language processing of regional Bengali languages remain limited. To bridge this gap, this work aims to investigate and document the characteristics of regional Bengali languages through comprehensive data-driven linguistic analyses, including phonetic and morphological studies. We also aim to study the feasibility of developing computational models, including Automatic Speech Recognition (ASR) systems, tailored to regional Bengali languages, which can facilitate applications like virtual voice command assistants and language processing tools. Our research findings will contribute to the understanding of regional Bengali languages, paving the way to foster the advancement of language technologies that can cater to the diverse linguistic needs of Bengali-speaking communities. Through this study, we intend to promote preservation of the regional dialects of the Bengali language, foster cultural inclusivity, and facilitate effective communication in the Bengali-speaking regions.
A conceptual framework for analyzing critical cactors of PDS learning experience and PDE employment experience
(BRAC University, 2023-12) Ahmed, Masum Uddin; Rahman, Mohammad Zahidur
This study aims to examine the relationship and status of education and employment for people with disabilities in Bangladesh. The study employs a mixed-methods approach, including a survey of people with disabilities to grasp their educational and employment situations. The purpose of this study investigation is to explore the factors impacting students and employees with disabilities in Bangladesh. The questionnaire consisted of a mix of numerical, categorical, and multiple-choice questions. This paper adopts multiple data science approaches to measure the reliability between survey items. Seven factors under three dimensions for students with disabilities (PDS) and eight factors under three dimensions for employees with disabilities (PDE) were examined to analyze the influence of their learning and employment experiences. A total of 208 responses were collected from students and employees, and 200 valid responses were retained after data cleaning. Necessary data pre-processing was applied. From the findings, eight factors influencing the learning experiences of students and employment experiences of employees were identified. Finally, the analysis results are presented in the form of suggestions for developing inclusive learning and employment opportunities for individuals with disabilities in Bangladesh. The survey reveals key obstacles for people with disabilities in Bangladesh, including accessibility issues, inadequate accommodations, negative attitudes, and undervaluation in education and employment. It underscores the urgent need for inclusive policies and more research to support their education and employment. The study highlights the requirement for diverse and more effective research methods to comprehend and provide support for individuals with disabilities in Bangladesh.
A data-driven institutional study of academic performance: exploring course, structure, and environment in higher education
(BRAC University, 2025-10) Shakil, Arif; Kazi, Sadia Hamid; Alam, Md. Golam Rabiul
Student performance reflects the intersection of academic, structural, and environmental factors rather than individual effort alone. This study conducts a largescale, data-driven analysis of institutional records from BRAC University to examine how these factors shape outcomes across time, delivery modes, and disciplines. Nine hypotheses were tested, encompassing the effects of online, hybrid, and inperson learning environments; the COVID-19 pandemic; course-level contributions to CGPA; prerequisite–core alignment; class size; Residential Semester (RS) contexts; high-failure course patterns; and inter-departmental performance differences. Using Pearson correlation, regression, ANOVA, Kruskal–Wallis, Welch’s t-tests, and mixed-effects models, the study identifies clear structural trends: RS participation strongly correlates with higher and more consistent GPA; class size shows weak, context-dependent effects; and persistent high-failure rates cluster in STEM gateway courses. Departments differ systematically—humanities and law programs maintain higher GPAs, while technical disciplines show greater variance due to assessment rigor and prerequisite dependency. These findings reveal that academic outcomes are institutionally patterned, not random. They underscore the need for data-informed curriculum design, departmental benchmarking, and early-risk intervention frameworks to promote equity, quality, and resilience in higher education.
A deep dive into node-level analysis with fusion RNN model for smart LTE network monitoring
(BRAC University, 2023-09) Islam, Md Rashidul; Alam, Golam Rabiul
Predicting and understanding traffic patterns have become important objectives for maintaining the Quality of Service (QoS) standard in network management. This change stems from analyzing the data usage on cellular internet networks. Cellular network optimiser frequently employ a variety of data traffic prediction algorithms for this reason. Traditional traffic projections are often made at the high-level or generously large regional cluster level and therefore has the lacking in precised forecation. Furthermore, it is difficult to obtain information on eNodeB-level utilisation with regard to traffic predictions. As a result, using the conventional approach causes user experience degradation or unnecessary network expansion. Developing a traffic forecasting model with the aid of multivariate feature inputs and deep learning techniques was one of the objective of this research. It deals with extensive 6.2 million real network time series LTE data traffic and other associated characteristics, including eNodeB-wise PRB utilisation. A cutting-edge fusion model based on Deep Learning algorithms is suggested. Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU) are three deep learning algorithms that when combined allow for eNodeB-level traffic forecasting and eNodeB-wise anticipated PRB utilisation.The proposed fusion model’s R2 score is 0.8034, outperforms the conventional state-if-the-art models. This study also proposed a unique method that thoroughly examines individual nodes for the Smart Network Monitor. This approach follows adjustments made to soft capacity parameters at the eNodeB level, aiming for immediate improvement or long-term network growth to meet a consistent QoS standard. The algorithm relies on expected PRB utilization.
A deep learning approach for pneumonia classiﬁcation from chest X-Ray images with ensemble modelling and explainable AI
(BRAC University, 6/8/2021) Akhter, Nasrin; Alam, Md. Ashraful
Pneumonia is one of those frightening diseases that has a high mortality rate among children and the elderly, with an estimated 2 million fatalities per year. Pneumonia aﬀects the poorest people in Africa and Asia the most, due to a lack of medical surveillance in such areas. It is responsible for 28 percent of all child fatalities in Bangladesh each year, and the number is likely to be considerably higher. In recent years, a number of computer-assisted diagnostic methods have been developed to assist in the detection of pneumonia. In this study, an eﬃcient model PNEXAI is proposed to identify pneumonia utilizing Chest X-Ray images. We gathered and classiﬁed data using VGG16, VGG19, ResNet 50, ResNet 101 and Inception v3. The accuracy rate of 97.17% was reached by VGG16, 97.69% by VGG19, 97.35%by ResNet50, 95.63% by ResNet101, and 94.86% by Inception V3, respectively. We then developed an ensemble model containing the top three classiﬁcations (VGG16, VGG19 and ResNet50) which delivered 98.46 % of best overall accuracy. Finally, to better comprehend our categorization, we included explainable artiﬁcial intelligence in our model.
A deep learning framework for arsenic skin disease detection leveraging dual-teacher knowledge distillation with a depthwise-separable convolution and KAN-based lightweight student model
(BRAC University, 2025-12) Mehedi, Md Humaion Kabir; Mridha, Muhammad Firoz
Arsenic poisoning in groundwater poses a major public health issue in Bangladesh. Chronic arsenic poisoning often leads to arsenicosis, a chronic disease characterized by cutaneous effects, including melanosis, leukocomelanosis and keratosis. Timely treatment of these lesions on the skin is important since prompt diagnosis and timely action are taken within the medical profession. However, this is challenging in rural areas. In most areas, there are no trained dermatologists, diagnoses tend to be subjective and there may be inadequate healthcare facilities. The proposed study involves the development of an explainable and lightweight deep learning framework for the automatic identification of skin diseases caused by arsenic, which will be computationally efficient and interpretable on a clinical scale. This system uses the dual-teacher knowledge distillation (KD) approach, in which two high-capacity models, InceptionV3 and Xception+InceptionM, are used to distill discriminative and contextual information to a small student model called Inception-Residual- KANNet (IR-KANNet). ArsenicSkinImageBD dataset was trained and validated using a stratified data split and 5-fold cross-validation to ensure class balance and model stability. The final evaluation found accuracy 0.9692, precision 0.9610, recall 0.9867 and F1-score 0.9737 on both infected and not-infected cases using stratified data split and a factor of 78.39% reduction in parameters over the teacher 1 and 35.29% in teacher 2 networks. These results indicate the appropriateness of the model for use in low-resource healthcare settings by providing convenient and reliable AI-based screening for arsenicosis detection. Grad-CAM++ and LIME make the model explainable and the resulting transparent heatmaps are consistent with the clinical regions in terms of lesions. This study is relevant for developing interpretable, efficient and domain-flexible medical AI models. It also provides a basis for further study of explainable knowledge distillation and edge-deployable solutions for the diagnosis of other dermatological disease
A document vectorization approach to Resume Ranking System(RRS)
(BRAC University, 2022-08) Nabi, Norun; Hossain, Dr. Muhammad Iqbal
Technology transformed the way how job seekers apply for a job and recruiter’s hunting for a precise pick . Now, paper version of resume already become an outdated version of job application method. Electronic resume replaces the old method thanks to its easier access to technology. When it comes to a particular job requirement, screening a rele vant resume among thousands is an exhaustive and time consuming recruitment process because the respective HR of an organization must have a proof read the entire resume set to select the right person in the right position, a key decision for any organization. Extracting the semantic meaning from resume is otherwise a daunting task. By making the selection process fast and accurate, organizations could save huge efforts and money. Using state-of-art-technology could be a way out. In the field of NLP, there are a range of tools to classify documents. Document vectorization technique is a huge popular one among tech-communities. Documents like resumes could be categorized and ranked by applying such techniques and tools. Therefore choosing a most suitable vectorization al gorithm is pivotal. It is aimed to build a custom trained model specialized in vocabulary of resume based on frequency based word2vec model such as TF-IDF. However, to compare between job descriptions and resumes, Cosine-Similarity is consid ered to be the primary algorithm to find matching resumes whereas k-nearest neighbor algorithm has been used to group the desired documents. But the limitation comes with using fixed vocabulary size. TOPSIS is the most popular among Multi Criteria Decision Making algorithms. Along with vector similarity score, Other parameters like years of experience, university rankings could be normalized to consider for final ranking score.
A graph mining-based approach to analyze the dynamics of the Twitter community of COVID-19 misinformation disseminators
(BRAC University, 2024-04) Hussna, Asma Ul; Alam, Md. Golam Rabiul
The abundant dissemination of misinformation on social networks has emerged as a worldwide threat, exerting an implicit influence on public opinion and endangering the progress of social, political, and public health domains in general. Amidst the rapid worldwide dissemination of the COVID-19 virus, unfortunately, misinformation about COVID-19 is being created and disseminated at a startling rate. The dissemination of misleading information has led to vast disorientation, social disruptions, and severe repercussions for health-related issues. Moreover, the dissemination of fake or misleading information via social media networking, particularly Twitter, during the COVID-19 pandemic has resulted in an extensive proliferation of information, commonly referred to as an “infodemic.” In order to combat the dissemination of fake news, we have proposed a research model that can predict fake news related to the COVID-19 issue on social media data using classical classification methods such as multinomial na¨ıve bayes classifiers, logistic regression classifiers, and support vector machine classifiers. In addition, we have applied a deep learning-based algorithm named DistilBERT to accurately predict fake COVID-19 news. These approaches have been used in this paper to compare which technique is much more convenient for accurately predicting fake news about COVID-19 on social media posts. The objective of this study is to understand how information is deviating and misinformation is spreading through social media during the COVID- 19 pandemic. Also, this research aims to examine the ecosystem of individuals who spread misinformation, with the objectives of comprehending their collective actions, identifying the most influential disseminators, and examining their online personas and profiles. We leverage the UUIG (User-User Interaction Graph) to capture the misinformation disseminators’ behavioral interactions. The following research analysis reveals the following significant findings: (a) the population of disseminators is growing rapidly even though today; (b) the community of disseminators comprises professional spreaders; above 3% of the fake news spreading population dominates others; and (c) they exhibit a high degree of collaboration among the fake news spreaders; we observe five big communities of collaborators. Our work represents a notable advancement in utilizing publicly available online data to gain insights into the community that spreads malicious misinformation about COVID-19.
A healthcare digital twin system based on blockchain technology
(BRAC University, 2023-04) Akash, Sadman Sakib; Ferdous, Dr. Md Sadek
Digital Twin (DT) is a technology that replicates any physical phenomenon from physical space to digital space in congruous with the physical form’s state. It does not confine to only spatial objects, any non-spatial scenarios can also be depicted with proper perception of the states. Though, DT technology was proposed with the incentive of revamping the intricate product lifecycle management in manufacture sector, other sectors like aviation, real states, healthcare, etc., have embraced it. By integrating DT in the healthcare sector, portrayal of patients in the digital space makes chances to create digital models, providing proper diagnosis, and evaluation facilities for digital healthcare services and Smart-health. However, determining a healthcare DT model for patient care and clinical purposes is seen as a challenging and imponderable task because of the lack of adequate data collection structures. Moreover, there are a number of problems in healthcare DT such as fragmented data and communication disorder which are making efforts futile. Also, the concept of healthcare DT is not formally defined and there lacks a consistent system architec ture and data model, using which the diverse data flow of a Healthcare DT can be perceived structurally and can be used for later purposes. The collected structured data and careful simulation with proper analysis can render propitious opportunities in grievous health situations. For this reason, formulating a comprehensive health care data model for Healthcare DT is a prominent and preemptory task. On the other hand, there are also security and privacy issues as healthcare data can be used in malicious ways. For these reasons, to acquire the codified and finesse data with total integrity and proper access control, blockchain can be incorporated with the DT technology. In this thesis, we present a mathematical concept data model to ac cumulate the patient relevant data in a structured and predefined way with proper delineation. Additionally, the provided data model is described in harmony with real life contexts. Then, we have used the patient centric mathematical data model to formally define the semantic and scope of Healthcare Digital Twin system based on Blockchain. Accordingly, the proposed system is described with all the key com ponents as well as proper protocol flow and evaluation. A short implementation of the proposed system has been conducted using Hyperledger Fabric and BigchainDB blockchains.
A hybrid deep learning model and explainable AI-based Bengali hate speech multi-label classification and interpretation
(BRAC University, 2022-09) Shakil, Mahmudul Hasan; Alam, Md.Golam Robiul
Data innovation has moved quickly in recent years, and various unfavorable alter ations have been made to the network medium. Social media platforms like Face book, Twitter, and Instagram are becoming more and more popular because they allow users to express their opinions through messages, photographs, and notes. In particular, in Bangladesh and other locations where the Bengali language is spoken. In any case, it has regrettably turned into a space with toxic remarks, cyberbully ing, and unidentified hazards. Numerous studies have been conducted in this area, but none have produced accurate results. Some effective pre-trained transformer models have been introduced. To identify Bengali malicious and non-malicious text at an early stage using simple Natural Language Processing (NLP). This study sug gests a Convolutional Neural Network with Bi-Directional Long Short-Term Memory (CNN-BiLSTM) hybrid strategy. This model can also classify any Bengali text data into six levels. Additionally, the transformed dataset is subjected to several conven tional Machine Learning methods using an estimator, and Explainable AI interprets these techniques (XAI). In the last stage, Stacking Classifier which is superior to any prior activity is used to ensemble all classifiers and the estimator.
A lightweight time-series analysis model through multi-teacher knowledge distillation for food price forecasting
(BRAC University, 2026-01) Zaman, Shifat; Alam, Md. Golam Rabiul
Accurate food price forecasting is critical for food security planning, particularly in developing nations like Bangladesh where price volatility can significantly im-pact vulnerable populations; however, existing deep learning approaches for time series forecasting often require substantial computational resources, limiting their deployment in resource-constrained environments. This study presents a novel multi-teacher knowledge distillation framework that transfers knowledge from an ensem- ble of three distinct teacher architectures—DLinear, PatchTST, and N-BEATS— trained on World Food Programme (WFP) Bangladesh commodity price data from the Dhaka Division to compact student models (MLP, GRU, KAN) through a multi- component distillation loss comprising prediction-level matching, feature-level alignment, and price-difference learning, with an uncertainty-weighted mechanism that focuses training on confident teacher predictions while dynamically weighting teacher contributions based on validation performance. Experimental evaluation on four food commodities (Lentils, Oil, Rice, and Wheat flour) with a 6-month input window demonstrates that the proposed approach achieves a Mean Absolute Error (MAE) of 1.959 BDT/unit with a Mean Absolute Percentage Error (MAPE) of only 3.73%, representing a 37% improvement over the supervised learning baseline, 69% improvement over traditional ARIMA, and 81% improvement over LSTM baselines, with the three-teacher ensemble distilled to an MLP student achieving the best results and outperforming all single-teacher and two-teacher configurations. The resulting student model requires only 200K parameters (compared to over 1M in the teacher ensemble) and achieves inference in sub-millisecond time on standard CPU hardware without GPU acceleration, enabling deployment in humanitarian field offices with limited computational infrastructure. This study contributes a reproducible, configuration-driven framework for knowledge distillation in time series forecasting, demonstrating that sophisticated ensemble-level accuracy can be achieved with lightweight models suitable for resource-constrained field deployment in food security applications.
A multi-level random key cryptosystem based on DNA encoding and state-changing mealy machine
(BRAC University, 2023-02) Taj, Towshik Anam; Hossain, Muhammad Iqbal
Cryptography allows our data to be transmitted without giving sensitive information away. This is the art of hiding the information from the malicious third party and make the data accessible to only the sender and the receiver. Building a complex cryptosystem has always been a challenge which can provide relentless security and is infeasible to break. This paper discusses about a hybrid cryptosystem which is inspired from the concepts of DNA cryptography and it is further strengthened using multiple components. A random key is used which is generated using run test of randomness, different DNA encoding combinations and a state changing random state mealy machine is used for further strengthening the security. This paper provides a detailed discussion regarding every components the authors used to build the system and also discusses about the combined system that has been built. This paper also discusses about the effectiveness and the performance of the proposed system to give an overview of its security measures and also provides some comparative analysis with existing works to back the claim on improved security features.
A semi-supervised federated learning approach leveraging pseudo-labeling for Knee Osteoarthritis severity detection
(BRAC University, 2024-06) Rifat, Rakib Hossain; Alam, Md. Golam Robiul
Within medical image analysis, appropriately classifying the extent of knee osteoarthritis is a significant obstacle, made more difficult by the scarcity of annotated data and strict privacy rules. Conventional approaches are hindered by the exorbitant expenses, limited availability of annotated datasets, as well as issues over the confidentiality of patient data. To overcome these challenges, we propose a method which is a Federated Learning Framework that utilizes pseudo-labeling we are calling it PLFL. Our innovative approach avoids the cost of human annotation and guarantees patient confidentiality through Federated Learning while reducing the dangers linked to adversarial assaults and annotation mistakes. Our proposed method works under the assumption that the server is the only custodian of gold label data, while the client side does not have any label data. The server utilizes gold-labeled data to train the global model and subsequently applies the federated learning approach. Clients add labels to unlabeled data by picking labels that meet or exceed a minimal threshold level of confidence in the prediction. Once data on the client side reaches the specified confidence score, it is added to the client’s dataset. Upon receiving the labeled data, the client initiates the training process and sends the weight of the local model. Subsequently, the server aggregates the weights of each model using the FedAvg technique. The thorough assessment of our system, in comparison to the standard client-server-based Federated Learning approach (CSFL) and FixMatchbased semi-supervised Federated Learning (FSSFL) approach, clearly shows significant performance improvements. Our framework PLFL showed superior performance compared to other explained techniques, with consistent accuracy, weighted average precision, recall, and an F1-score of 0.88. Significantly, it outperforms both CSFL and FSSFL Frameworks, significantly enhancing model performance and efficiency. The proposed framework achieves an accuracy of 93.07% for the healthy class, 64.00% for the moderate class, and 100% for the severe class. Furthermore, our system has exceptional prediction precision, especially in detecting moderate and severe instances of osteoarthritis, surpassing rival frameworks. This is seen in the notable progress in accurately forecasting moderate and severe categories, highlighting the effectiveness of our method. The pseudo-labeling-based framework had the shortest duration for label generation and model training, 3.2 times shorter than the best-performing model of the traditional Federated Learning Framework (CSFL) and 1.7 times lower than the best-performing model of the FixMatch-Based Federated Learning Framework (FSSFL). This thesis proposes an innovative investigation into identifying knee osteoarthritis severity, the first instance of applying semi-supervised and federated learning approaches in this field. Our goal is to stimulate progress in medical image analysis by using our innovative technique, resulting in more precise diagnoses and better patient outcomes.
A squeeze and excitation ResNeXt-based deep learning model for Bangla handwritten basic to compound character recognition
(BRAC University, 2021-08) Khan, Mohammad Meraj; Rahman, Md. Khalilur
With the recent advancement in artificial intelligence, the demand for handwrit- ten character recognition increases day by day due to its widespread applications in diverse real-life situations. As Bangla is the world’s 7th most spoken language, hence the Bangla handwritten character recognition is demanding. In Bangla, there are basic characters, numerals, and compound characters. Character identicalness, curviness, size and writing pattern variations, lots of angles, and diversity makes the Bangla handwritten character recognition task very challenging. There are few papers published recently which works both Bangla numeral, basic and compound handwritten characters, but the accuracy level in all three areas is not so satisfac- tory. The main objective of this paper is to propose a novel model which performs equally outstanding in all three different character types and to increase the effi- ciency to build a real-world Bangla Handwritten character recognition system. In this work, we describe a novel method of recognition for Bangla basic to compound character using a very special deep convolutional neural network model known as Squeeze-and-Excitation ResNext. The architectural novelty of our model is to in- troduce the Squeeze and Excitation (SE) Block, a very simple mathematical block with simple computation but very effective in finding complex features. We obtained 99.80% accuracy from a bench-mark dataset of Bangla handwritten basic, numer- als, and compound characters containing 160,000 samples. Additionally, our model demonstrates outperforming results compared to other state-of-the-art models
A sustainable Bitcoin architecture
(BRAC University, 2022-04) Monem, Maruf; Alam, Md. Golam Rabiul
Cryptocurrencies are the new form of trade that has revolutionized how we look into our financial institutions. Bitcoin dominates the industry with the highest market share among the hundreds of other cryptocurrencies. However, high energy consumption leading to increasing carbon emission, prioritizing high-value transactions, and long waiting times are some of the flaws preventing it from reaching its full potential. Due to the block rewards getting halved every four years, miners and researchers are fearful that this would be the breaking point of Bitcoin’s success. One of the ways to tackle and hopefully reduce this problem while bringing wider adaptability is by ensuring faster transactions. Currently, Bitcoin has an average block size of 1MB, which many researchers and enthusiasts believe is insufficient. To tackle these limitations, we have proposed two different ideas. Our first concept proposes an industry 4.0 compliant next-generation Bitcoin architecture by introducing a dynamic and sustainable block concept. Using our improved knapsack algorithm, a priority-based 0/1 knapsack and advanced priority-based 0/1 knapsack, we can ensure a balanced transaction selection, quicker verification, higher transaction throughput, reduced carbon emission, and increased earnings for the miners. Moreover, with the addition of only one of our proposed sustainable blocks, we can cut down verification times by 50% and increase throughput by 2.56 times. We can also reduce carbon emissions per transaction by 62.318%, which would help reduce Bitcoins’ large carbon footprint, enabling us to approach greener digital transactions. In the second concept, we further try to improve the block sizes using the help of machine learning and artificial intelligence. Our proposed model analyzes the network’s activity, such as incoming transaction frequency and other aspects, to adjust block sizes. The model can predict block sizes with 61.12% accuracy, and we can see a positive change in the amount of fees earned by miners (9.3%), transaction count and transaction per second (66.75%). With the help of our model, Bitcoin would be able to dynamically change the block size based on the transaction activity, resulting in shorter wait times, thus increasing wider adaptability and sustainability.
A transformer based approach to detect the sentiment of drivers in ride sharing platforms
(BRAC University, 2024-04) Chakraborty, Sovon; Sadeque, Farig Yousuf
Globally, ride-sharing is very popular, especially in developed countries. The scheme has been launched in many developing countries, and Bangladesh is no exception. The ongoing transportation problem and traffic jams make this country vulnerable economically. The impact of COVID-19 has snatched the jobs of many people. The ride-sharing platform allowed them to grab a chance to be self-dependent. On the contrary, the increasing hike in daily vehicle accessories, fuels, and parts makes it difficult for a rider to earn his bread and butter. In this research, the author focuses on the impact of ride-sharing and drivers on the Bangladeshi economy. Along with this, many social and economic statuses are analyzed. At first, a dataset was prepared after discussing it with 2234 drivers. Extensive exploratory data analysis was performed to find insightful information from the dataset. Later, the dataset is preprocessed precisely before feeding into numerous Machine Learning and Deep Learning architectures. A comment from each of the riders is also taken to understand the sentiment of these riders. Three sentiments have been considered, namely Positive, Negative, and Neutral. The researchers have adopted an optimized BERT transformer-based approach to validate the dataset and classify Bengali comments correctly. The model can outperform the state-of-the-art architectures in numerous performance metrics. The optimized model shows a 80.63% F1-score in the training dataset, whereas it shows an 84.53% F1-score in the validation set. Finally, the black box model is interpreted with the aid of Explainable Artificial Intelligence.
A viseme recognition system using lip curvature and neural networks to detect Bangla vowels
(BRAC University, 2016) Akhter, Nahid; Chakrabarty, Dr. Amitabha
Automatic Speech Recognition plays an important role in human-computer interaction, which can be applied in various vital applications like crime-fighting and helping the hearing-impaired. It consists of two domains – Audio Speech Recognition and Visual Speech Recognition. This thesis is based on Recognition of Speech in the visual domain only, i.e. it involves recognizing speech without the presence or support of any auditory signal. So far, a lot of research has been done on lip-reading in English and some amount on French and Chinese, as well as few other languages, but not much research has been done on lip-reading in Bengali. This thesis work provides a new approach to lip reading Bengali vowels using a combination of the curvature of the inner and outer lips and Neural Networks. The method uses a more robust and faster algorithm to detect the lip contour than conventional methods used so far, such as Active Contour Model, Active Appearance Model and Active Shape Models. The method used for feature extraction is also new. It makes use of coefficients of the curves of the inner and outer lips. This way, it makes use of a lesser number of parameters to represent the shape of the lip when pronouncing a vowel. Moreover, the method is also robust to alignment of lips at different angles and can work with low resolution pictures also. Finally, for recognition of the viseme, a Backpropagation Neural Network is trained and simulated using gradient descent method.
Advanced video analytic system for posture and activity recognition: leveraging MediaPipe, CNN-LSTM, and ensemble learning for fall and unstable motion detection
(BRAC University, 2024-10) Siraj, Farhan Md.; Zereen, Aniqua Nusrat
Human posture detection and classification are vital in monitoring activities, especially in health and safety contexts, such as fall detection in elderly care. This thesis presents a comparative study of two machine learning approaches for real-time human posture calssification using real time video data, a traditional feature-based approach using a Voting Classifier, and a deep learning appraoch utilizing a Convolutional Neural Network-Long_Short-Term memory (CNN-LSTM) model. The feature-based method incorporates pose estimation using MediaPipe to extract human body landmarks, followed by classification using an ensemble of Rnadom Forest, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). On the other hand, the CNN-LSTM model captures both spatial and temporal dynamics of video sequences by extracting visual features through Convolutional Neural Network (CNN) and modeling temporal dependencies via LSTM. The models are evaluated on a dataset for four postures-Fall, SIt, Stand and Unstable-with promising result. This work demonstrate the effectiveness of combining pose-based features with voting classifiers and the power of deep learning in sequential data, offering a robust solution for real-time posture classification systems.