Thesis (Bachelor of Science in Computer Science)

Browse

Author: search.filters.author.Rahman, NaimurStart date: 2020

Search Results

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Optimizing abstractive summarization with fine-tuned PEGASUS
    (BRAC University, 2023-09) Rafi, Sadiul Arefin; Rahman, Naimur; Islam, Kazi Nazibul; Ahmad, Ha-mim; Sadeque, Farig Yousuf
    Abstractive text summarization is the technique of generating a short and concise summary comprising the salient ideas of a source text without making a subset of the salient sentences from the source text. The introduction of transformer models such as BART, T5, and PEGASUS has made this sort of summarization process more efficient and accurate. The objective of this paper is to analyze the performance of different transformer models, compare them to find an efficient model and fine-tune the model on csebuetnlp/xlsum English corpus. The performance of the generated summaries from the fine-tuned PEGASUS models is evaluated using the ROUGE metric, which basically compares the auto-generated summaries with human-created summaries. The fine-tuned PEGASUS model gives a state-of-the-art performance on the XLSum English Corpus.
  • Thumbnail Image
    Item
    Leveraging sequential deep learning models for detecting multitude of human action categories
    (BRAC University, 2023-09) Pranta, Kazi Al Refat; Islam, Fahad Mohammad Rejwanul; Ahmed, Khandakar Fahim; Saha, Prince; Rahman, Naimur; Reza, Tanzim; Rahman, Rafeed
    In today’s world, where science and technology are constantly evolving day by day, people are drawn to tangible experiences and visual representations. There’s a growing effort to teach machines about human movements and postures to enable smart decision-making. This has led to increased interest in the field of human action recognition (HAR) among researchers globally. Our research focuses on implementing advanced technologies to address criminal activities, specifically emphasizing Human Activity Recognition (HAR). Moreover, our dataset includes 1275 videos, covering 20 different actions involving both violent and non-violent behaviors. In addition, we have developed a pipeline that utilizes YOLO-v8 to extract background, followed by models for accurate video classification. two models,conv-lstm and lrcn, were incorporated into our deep learning pipeline. Through our observations, we found that the LRCN model outperformed the other model, achieving an accuracy of 62% and an F1 score of 60% for the 20 classes, for 17 classes an accuracy of 63% and an F1 score of 66%. for binary classification LRCN got accuracy of 88% and an F1 score of 87%Our research focusses the potential of advanced technologies to significantly improve Human Activity Recognition (HAR) in addressing various aspects of criminal activities in real-time scenario. This marks a substantial step forward in intelligent decision-making and public safety.
  • Thumbnail Image
    Item
    Detection of prodromal parkinson’s disease with fMRI data and deep neural network approaches
    (BRAC University, 2021-06) Shahriar, Farhan; Dey, Amarttya Prasad; Rahman, Naimur; Tasnim, Zarin; Tanvir, Mohammad Zubayer; Parvez, Zavid
    Parkinson’s Disease is the second most common neurological disease after Alzheimer’s Disease. The disease is incurable. However, if the disease can be detected earlier, then the consequences of it’s effect can be relieved. The early phase of PD is called by Prodromal Parkinson’s Disease. The symptoms of the Prodromal phase includes hyposmia, constipation, mood disorders, REM sleep behavior disorder, olfaction dis orders etc. RBD or REM sleep behavior disorder is the most common symptom of Prodromal PD. In this study, we used various deep convolutional neural network architectures and trained them to detect Prodromal PD patients. We collected 20 Prodromal patients and 20 healthy control subject data from the PPMI website and applied CNN architecture mobilenet v1, incception v3, vgg19 and inception resnet v2 to achieve our goal. We ensembled inception resnet v2 and mobilenet v1 with the hope of getting a better result as well. However, we successfully carried out our training and with mobilenet v1 we gained the highest classification accuracy of 81.22%. Inception resnet V2, inception v3, vgg19 and ensemble model achieved respectively 75.30%, 62.55% and 63.32% accuracy.