Thesis (Bachelor of Science in Computer Science)

Browse

Has files: YesSubject: search.filters.subject.Deep learning

Search Results

Now showing 1 - 10 of 150
  • Thumbnail Image
    Item
    A universal photography suggestion system utilizing composition detection, orientation detection, and subject position detection
    (BRAC University, 2025-06) Niloy, Iftikhar Shams; Proma, Syeda Mahjabin; Dofadar, Dibyo Fabian; Ahmed, Md. Sabbir
    Photography is one of the most popular hobby and images are one of the most important content types on social media, and the impact of a photo often hinges on its composition as much as its subject. In response to this, we proposed a system that classifies the compositional structure, detects orientation and subject of a given photo and suggests improvements based on established photography rules. For the classification of the composition, the photo will be categorized into one of five classes(CC, ROT, LL, FIF, PAT). Then, it will determine the orientation of an image. Lastly, this system uses YOLOv8 object detection model to find the objects of a photograph and through logics and conditions the subject is determined. The proposed system will provide the final suggestion based on the three results of the three proposed models. The main goal of the research is to develop a suggestion system that utilizes the detection models built using Deep Learning(DL) algorithms and find the optimal models that will accurately determine the composition, orientation and subject (if any) of a photograph. We have achieved up to 74.34% accuracy in our composition detection model and a minimum of 0.5870 mean square error (MSE) on our orientation detection model. The subject detection conditions capable of properly detecting the subject of an image most of the cases. Our approach aims to assist users in improving their photography skills and elevating the quality of visual content on any media platforms.
  • Thumbnail Image
    Item
    Automated hazard detection for AR/VR Mars terrain navigation using computer vision
    (BRAC University, 2026-01) Rhidy, Tasin Ahsan; Rahman, MD Touhidur; Shuvo, Istiak Zaman; Mahmood, Saiyed Mubasshir; Rafi, Abrar Mojahid; Alam, Md. Ashraful; Alam, Md. Golam Rabiul; Tasnim, Sanjida
    The exploration of Mars brings about a series of challenges that are occasioned by the risky topographical features, random weather patterns, and the fundamental need to have self-driving equipment. The study builds a combined computer vision and immersive technology system to improve the safety of the human astronauts and robot rovers in their navigation on the surfaces of the Martian environment. Our solution is a multi-modal deep-learning system consisting of object detection, semantic segmentation, and monocular depth estimation to generate complete hazard awareness in simulated Mars environments. We use datasets to train terrain classification models that are able to detect important surface features such as rocks, boulders, and potholes as well as other geological features. The system combines a number of deep-learning networks to detect hazards in real-time and locate bounding-boxes, semantic-segmentation, and pixel-level terrain-classification as well as a depth-estimation architecture to give the system spatial information of the Martian terrain. These models are synergistically used to produce an environmental cognition that drives into an AR/VR interface that provides users with visual cues in safe path planning. The AR/VR element converts raw computer-vision data into usable navigation data, and deciphers warnings of hazards and terrain complexity data to the Martian landscape. The initial studies have shown strong detection of varied terrain conditions, and the multi-modal strategy has a great benefit on improving the safety of navigation in comparison to the single-modality systems. The study has been applied to the development of autonomous planetary exploration technologies and created a scalable model of pre-mission astronaut training and rover operation plan.
  • Thumbnail Image
    Item
    A 3D convolutional neural network architecture for early detection of coronary artery blockage (Coronary-3D-UNet)
    (BRAC University, 2026-01) Sazid, Ahanaf Abid; Rahman, Mushfiqur; Akon, Md. Sabbir; Sadik, Jauad Ahmed; Chowdhury, Md. Jabed; Alam, Md. Ashraful
    Coronary Artery Disease (CAD) is a major cause of death all over the globe where the estimated number of annual deaths are 9 million. Early and precise diagnosis of the stenosis of the coronary arteries is critical in providing clinical intervention and better patient outcomes. Despite the fact that invasive coronary angiography (ICA) is regarded as the most effective diagnostic tool, the procedure has certain risks and is not convenient enough to be used widely because of its inapplicability to whole population screening. Coronary Computed Tomography Angiography (CCTA) is a non-invasive, high-resolution alternative but manual analysis of CCTA images is time consuming and subject to inter-observer error. This paper presents Coronary-3D-Unet, a parameter-efficient 3D convolutional neural network to perform automated stenosis assessment and coronary artery segmentation on CCTA volumes. The residual learning and dual attention mechanisms (spatial and channel) and multiscale feature fusion are incorporated into the proposed architecture to improve vessel representation and resistance to various challenges such as small vessel structures, low contrast, and imaging artifacts. The framework allows automatic stenosis grading of severity as mild, moderate, and severe geometrically. On a publicly available benchmark dataset, experimental results demonstrate that Coronary-3D-Unet reaches a mean Dice Similarity Coefficient (DSC) of 76.6% and bests the official ImageCAS baseline with a significantly lower model complexity. The model is based on an efficient number of 3.8 million parameters, which is 83% less than the conventional dense 3D networks, so the model can be used to infer efficiently and be deployed practically. The model can also be integrated with clinical Picture Archiving and Communication Systems (PACS) that facilitate real-time analyses and better clinical usability.
  • Thumbnail Image
    Item
    A data-driven exploration of Stratospheric Ozone dynamics : Bridging regional Ozone insights with environmental policy
    (BRAC University, 2025-12) Hossain, Md. Abir; Nawshin, Sadia; Rahman, Sabira; Abdullah-Al Saud, Shah Md.; Rubaia, Saba; Ahmed, Md. Sabbir
    The stratospheric ozone can be very crucial in protecting the Earth against harmful UV radiation, as well as its restoration after the 1987 Montreal protocol, is unevenly spread in various regions. This research is a data-based examination of the longterm dynamics of the ozone in various countries despite having different climatic and geographical settings, which showed specific recovery in various regions. The study models complex seasonal and nonlinear ozone behavior, using the support of more complex feature engineering, which incorporates lag variables, rolling averages, and temporal indicators based on advanced deep-learning models: LSTM, GRU, TCN, Transformer, and hybrid solutions. Model assessment, which is based on the combination of accuracy measures and uncertainty estimation, indicates that LSTM is the best in terms of explanatory performance, and GRU achieves the lowest in terms of MAE and RMSE in all six climatic regions. Another meta-analysis conducted across all regions further synthesizes recovery slopes and prediction error and levels of uncertainty, with strong recovery rates in the Tropical, Temperate regions, and slower or more erratic rates in Polar, Subpolar, and Arid regions. A policy modeling framework based on the use of data-driven insights to inform climate-aligned policies in SDG 13 (Climate Action), the mitigation of UV-risks in SDG 3 (Good Health and Well-being), and the improvement of environmental planning in SDG 11 (Sustainable Cities and Communities) is also introduced in the study. This framework offers an evidence-based and scalable policy instrument to monitor the environment in the long term and make decisions to bridge long-term ozone recovery and policy action recommendations to sustainable climate decisions.
  • Thumbnail Image
    Item
    Adaptive traffic signal control for urban intersections using reinforcement learning: a SUMO simulation case study on Gulshan-2
    (BRAC University, 2025-02) Shamim, MD Shahadat Hossain; Chowdhury, Tawhid; Deep, Sadid Arman; Bhuban, Riazul Hoque; Chakrabarty, Amitabha
    Traffic congestion is an ever-growing concern in rapidly increasing population and increase of vehicles on roads. On top of that, our road condition, pedestrian be- havior, and driving behavior makes the situation much worse. As Bangladesh has started adapting to the fixed-time traffic light system, finding the optimal adaptive traffic light management system is required to mitigate the issue. Keeping the lack of intra-vehicular communication devices in mind, we have studied different traffic light controlling systems and how these perform on Bangladeshi roadways. In our study, we have used traffic simulator to simulate Gulshan-2 intersection, one of the major congestion points in Dhaka city. Based on traffic data, we have created a vehicular network system to find out how the existing traffic light models per- form on mitigating the traffic jam using an adaptive traffic management system. In our study, we simulate the Gulshan-2 intersection using a realistic traffic dataset de- rived from field data and analyze three control models, Fixed-Time, Q-Learning, and Deep Q-Network (DQN). The results show that in ideal traffic condition, when no anomalies are present; the Q-Learning controller reduced average vehicle delay by approximately 60-70% and improved throughput by 35-45% compared to the fixed- time system, demonstrating its superior adaptability to dynamic traffic flows. How- ever, when we add real-life anomalies such as potholes and jaywalking, the baseline controllers’ performance drops. By acknowledging the drawbacks of QL controller in abnormal scenarios, we have introduced a hybrid model CRQL (Congestion Re- sponsive Queue Learning) and made some improvements. In abnormal situation CRQL performs better in regard to throughput, which is a 76.9% improvement over Fixed-Time, 37.4% improvement over Q-Learning and 215.2% improvement over DQN. This reflects a overall performance improvement under disruptive and non- stationary environment. In short, the study helps to see how existing models perform under real-life anoma- lies, which factors are significant to create Ad hoc model for our extreme pedestrian patterns and road conditions and how our proposed solution; the CRQL ad hoc or hybrid model performs to address the problems.
  • Thumbnail Image
    Item
    An efficient deep neural architecture for detecting different stages of age-related macular degeneration
    (BRAC University, 2026-01) Rafi, Iftekher Uddin; Shamiha, Antara; Akanda, Md. Ibrahim; Saha, Richa; Ahmed, Md Sabbir; Alam, Md. Ashraful
    Age-related macular degeneration (AMD) is a leading cause of vision loss in the elderly population, characterized by progressive retinal degeneration that advances through distinct stages: early, intermediate, and late (advanced). Clinical management, early intervention, and prevention of irreversible visual impairment require proper and timely diagnosis of these stages. Nevertheless, the manual interpretation of retinal images is still laborious and prone to interobserver error, which demonstrates the necessity of the use of high-quality automated diagnostic systems. Our work in this paper suggests an effective deep neural network to identify and classify the stages of AMD with the use of retinal imaging, namely, optical coherence tomography (OCT). The research first performs a binary detection to make the distinction between AMD and Non-AMD cases. We then categorise the cases of detected AMD on the basis of early, intermediate and late stages. The framework proposed makes use of transfer learning using various state-of-the-art convolutional neural network (CNN) backbones, such as ResNet50, EfficientNetB4, ConvNeXt Small, ConvNeXt Tiny, and EfficientNetV2_L that have been selected due to their effective representational ability and efficient computing. These are fine-tuned models that are used to observe subtle pathological changes related to drusen accumulation, pigmentary defects, geographic atrophy, and neovascularization throughout the disease range. The architecture assumes the combination of the standardized preprocessing, efficient training methods, and the relative assessment of the chosen networks to find an effective and resource-efficient model of the multi-class AMD staging. Experimental findings demonstrate good performance. ConvNeXt Small reaches 97.9% accuracy in detection and 83.5% in staging. It outperforms ResNet50 (97.5% detection, 77.4% staging), EfficientNetB4 (97.5% detection, 69.9% staging), ConvNeXt Tiny (96.7% detection, 78.2% staging), and EfficientNetV2_L (96.7% detection, 80.4% staging). All models improve early-stage sensitivity, with recalls above 86%. Experimental results demonstrate that the proposed approach achieves robust performance across all disease stages, with improved sensitivity in early-stage detection. This work highlights the potential of modern deep learning models as scalable decision-support tools for large-scale screening and clinical diagnosis, contributing to improved outcomes in the management of age-related macular degeneration.
  • Thumbnail Image
    Item
    A deep learning-based approach to detect Parkinson’s disease through speech analysis
    (BRAC University, 2026-01) Ovi, Anjon Biswas; Rifat, Jahidul Hassan; Azim, Mashrur; Mahi, Fatin Ishraq; Hossain, Md Arafat; Siddiqui, Md. Saiful Bari
    Parkinson’s Disease (PD) is a progressive type of neurodegenerative disorder in which timely diagnosis is extremely important in managing the disease, yet it is difficult because of the insidious nature of the symptoms. The conventional diagnostic approaches usually use subjective clinical evaluation which might fail to reveal early biomarkers. This study fills this gap by proposing a deep learning-based framework that can be used to detect PD non-invasively with speech signals. In this paper, we present a multi-model fusion architecture that makes use of two different representations of one audio signal, which are Mel-spectrograms to represent spectral-visual features and Wav2Vec 2.0 embeddings to represent high-level contextual phonetic features. We extractedWav2Vec2 embedding features from a fully connected neural network that was trained on the Wav2Vec2 embeddings. We then ran the melspectrograms though CNN and other convolutional networks such as ResNet18, ResNet50, Inception v3, Efficient Net B0, and Efficient Net B4. We then extracted melspectrogram features from all these convolutional networks and concatenated them seperately with the Wav2Vec2 embedding features. The cocnatenated features then went through a commonly shared final classifier. One of the most important contributions of this work is the introduction of a new rational sine activation function that gives the model a better capability of learning nonlinear decision boundaries. The results of the experiment conducted on the Italian Parkinson voice and speech data show that the proposed fusion model attains a test accuracy of 85.29% which is higher than what the other architectures that involve transfer learning models such as ResNet18, ResNet50, Inception v3, Efficient Net B0, and Efficient Net B4 have achieved. Moreover, a strict cross-corpus test on the unseen Vowel dataset shows that although domain shift is a universal problem, the suggested fusion architecture demonstrates better generalization with respect to the baseline models, which are adversely affected by the problem. The paper has validated the hypothesis that a combination of mixed signal representations with mathematically optimized activation functions is a promising avenue towards automated and objective screening of PD.
  • Thumbnail Image
    Item
    A coarse-to-fine hierarchical framework for bone marrow cell recognition: integrating morphological features with class-specific augmentation and multi-perspective explainable AI
    (BRAC University, 2025-10) Mamun, Abdullah AL; Haider, Zarin Tasnim; Montaha, Sidratul; Jahan, Ismat; Mukta, Jannatun Noor; Nasim, Hamim Ibne
    The classification of bone marrow (BM) cell is essential for the diagnosis of many haematological disorders. Automated cytological analysis still suffers from extreme class imbalance, very high morphological similarity between cell types, and poor interpretability of most artificial intelligence (AI) models, despite advances in medical imaging and deep learning. We present an interpretable, state-of-the-art framework for BM cell classification based on a large-scale dataset with 171,374 single-cell images annotated by experts with 21 classes. Given the high severity of the class imbalance (originally 3678:1 at times), we created a new subset of 95,865 images from the overall dataset through segmentation and feature extraction. Through this process, overrepresented classes were down-sampled, while specific types of augmentation were applied to underrepresented classes to restore balance, resulting in a ratio of 1435:1. Our recursive segmentation approach, based on CMYK (Cyan, Magenta, Yellow, and Black) and HLS (Hue, Saturation, and Lightness) colour spaces, reliably identifies nucleus, cytoplasm, and whole-cell regions. From these areas, we processed 144 biologically motivated shape, colour, texture, and fractal attributes. We build a hierarchical two-stage classification model named HierEff-S2, where an EfficientNet-B4 backbone assigns each cell to one of six morphological groups. Then, group-specific EfficientNet-B3 models perform fine-grained classification within each group. With 21 classes, this architecture obtains 86.1% accuracy and outperforms other models, including VisionMamba, Ensemble Model, and MobileNet. To promote clinical interpretability, we combine two explainable AI methods to visually highlight cell regions that lead to the model predictions: Grad-CAM and LIME. Using XAI, we report 95.2% correctness at the image level, thus providing biologically meaningful attention to the model.
  • Thumbnail Image
    Item
    Identification of IoT vulnerabilities using DRL
    (BRAC University, 2020-10) Bordhen, Joy; Islam, Tarikul; Sakib, Tahmid Al; Joy, Tanvir Ahmed; Tasnim, Sara; Hossain, Muhammad Iqbal
    Security issues have been a threat to our modern life even after all the inventions in modern data networks. It has come to a stage where with great innovation comes greater risk. Recently, security issues in IoT devices are increasing day by day with the vast uses of IoT devices in our everyday life. In today's world, the most valuable thing is information, the more information you possess the richer and more powerful one is. Due to the increase in use, the online attackers have been recently targeting IoT devices to gain monetary benefit or acquiring different sensitive information from the crucial sources. Maximum IoT devices are very simple to get hands on illegally. Since IoT devices are not well equipped with proper security measures in order to keep them easily accessible and low cost, the hackers are easily accessing these devices. However, computer and digital devices often appear to become more unstable and vulnerable to malfunctions and vulnerabilities due to cyber attacks. Also, this is economically harmful which is quite frustrating. Securing these have been quite an important issue for many years. Thus, to make sure the information is safe and to overcome these vulnerabilities different measures are taking place to minimize these attacks. Several surveys are put forward to address these IoT based topics including intrusion detection, threat minimization and prevention of these attacks. It is not possible for the users to keep track of the intrusion every time manually. So an autonomous security measure should be taken to prevent the intrusion. In order to do this we are using Deep Reinforcement Learning technique to detect this intrusion and prevent them with the existing data and model as it is highly efficient in tackling complex, diverse and, in particular defense of highextent cyber attacks. Reinforcement learning (RL) operates in a rotation of sense action goals. Since reinforcement learning acquires information directly from the environment, it is distinct from supervised learning that learns from the examples given. We are trying to detect and prevent malware attacks like viruses, ransomware and everything. For that, patterns need to be found by using deep learning or deep reinforcement learning algorithms. Furthermore, the system will be analyzing multiple attacks from systems that are already infected to do so and finally we will be coming up with a pattern created out of data analysis to prevent further attacks.
  • Thumbnail Image
    Item
    Pneumonia detection and classification using neural network
    (BRAC University, 2020-10) Showkat, Jobayer Bin; Alavi, Salsabil Hossain; Sadman, Sean; Ajwad, Rasif
    Medical image classification is of vital importance when it comes to clinical treatment. Numerous imaging techniques exist in diagnosis of various diseases, but X-rays remain one of the most popular techniques. Given that, X-rays are inexpensive and the easiest to perform, it is one of the most frequently used radiology examinations for diagnosis. X-rays are performed in different parts of the body. In this paper, we will focus on chest X-rays which give us images of the lungs, heart, airways and the bones which are present in the chest and spine. Chest X-rays can also display fluid inside the lungs or the area surrounding the lungs. The image produced by chest X-rays can help doctors determine many lung-diseases. However, examining chest X-rays clinically can be tedious and complex. Therefore, computer aided detection can help to achieve more accurate and simpler ways of acquiring correct diagnosis. Computer aided detection techniques include the use of machine learning algorithms and deep learning methods.