Thesis (Bachelor of Science in Computer Science)

Browse

Has files: Yes

Search Results

Now showing 1 - 10 of 868
  • Thumbnail Image
    Item
    Spatiotemporal analysis of air pollution using advanced machine learning techniques
    (BRAC University, 2026) Rahman, Shafin; Islam, Naeem; Rahman, Md. Shoaibur; Hridoy, Md. Moniruzzaman; Alam, Md. Ahasanul
    This thesis presents a unified framework for spatiotemporal analysis of air pollu- tion using advanced machine learning to enable short-horizon, citylevel forecasting and operational decision support. A leakage-safe, multi-source dataset is curated for 20 cities across Bangladesh and China, integrating pollutant observations with spatiotemporal covariates (e.g., meteorological and contextual signals) to model ur- ban pollution dynamics under heterogeneous conditions. The forecasting task is formulated as multi-output time-series regression over PM2.5, PM10, NO2, SO2, and CO. To capture short-term fluctuations and longer temporal dependencies while exploiting cross-pollutant structure, a multitask CNN–LSTM architecture is de- veloped with a shared feature backbone and pollutant-specific prediction heads. Performance is benchmarked against classical machine-learning baselines (including Random Forest and XGBoost) under cityaware evaluation to assess both accuracy and robustness. To address regional data imbalance, a cross-country transfer learn- ing strategy is evaluated by leveraging representations learned from data-rich source cities to improve forecasting in data-scarce target cities. Forecast reliability is en- hanced via Monte Carlo Dropout to estimate predictive uncertainty, while SHAP and Integrated Gradients provide complementary explanations of feature influence and temporal attribution. Finally, an early-warning episode detection layer converts forecasts into event-oriented alerts and diagnostics to support practical monitoring workflows. Overall, the proposed pipeline delivers more accurate, uncertainty-aware, and interpretable multi-pollutant forecasts suitable for risk-sensitive air quality man- agement in heterogeneous urban environments.
  • Thumbnail Image
    Item
    Culturally adaptive neural network for detecting cybersecurity vulnerabilities in Bangladeshi web applications
    (BRAC University, 2026) Shaolin, Mohosina; Nawar, Fariha; Siddique, Arik Ahmed; Shams, Shaikh Mohammad Ali; Maliyat, Nafisa; Mostakim, Moin.
    As cyber threats become more complex and frequent, conventional methods for detecting website vulnerabilities, such as rule-based and heuristic approaches, faces significant difficulties, including limited adaptability, high rates of false positives, and a lack of contextual insight. This study presents a predictive model based on neural networks aimed to actively evaluating website security. By applying essential features like security headers, SSL/TLS settings, and SQL injection vulnerabilities, the model detects complex patterns and irregularities, enabling precise identification of emerging threats and vulnerabilities. This approach uses data-driven feature engineering and training with custom neural architectures, for comparison we used random forest and gradient boosting, For explainability we used SHAP followed by evaluation metrics such as precision, recall, and F1-score. Key results show improved accuracy, reduction of false positives, automated monitoring of configurations, and enhancement of resilience against adversarial attacks. Although neural networks show significant potential for transformation, challenges related to transparency, computational demands, and data imbalance are acknowledged. This highlights the necessity for ongoing learning, scalability, and integration with current frameworks, laying the groundwork for robust and adaptable web security strategies.
  • Thumbnail Image
    Item
    An end-to-end framework for anomaly detection and categorization
    (BRAC University, 2025) Islam, MD. Farhan; Islam, Rehnuma; Reza, Syed Rahin; Tasnim, Saifa; Nipu, Anipa Akter; Rahman, Rafeed
    In this study, we proposed an end-to-end framework for anomaly detection, classification in Industry 4.0 using deep learning models YOLO V8 and ResNet on the MVTec Anomaly Detection(MVTec AD) dataset. The framework is based on defect detection, anomaly localization. The multitask queues in YOLO V8 guarantee both: fast and precise detection in real time, while ResNet primarily suited for classification, complete with top notch precision and recall metrics. The metrics used for evaluation (including AUC, accuracy, precision, recall, F1 score and AP) confirm the good performance of the models. We also provide decision surface visualizations through Grad-CAM and Integrated Gradients that will help you understand some of the decisions made by the model. The YOLO V8 performed optimal on real-time detection tasks and ResNet performed best on classification accuracy, as highlighted through the results. This framework allows for the automation of anomaly detection and the resolution through investigation, unlocking future opportunities for real time anomaly detection and management.
  • 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
    Real-time aviation anomaly detection and multi-label classification using deep learning on multivariate sensor data
    (BRAC University, 2026) Yeasin, Sakib Rayhan; Rahat, Md. Atik Hasan; Nakib, Shafaat Jamil; Mitra, Debjoty; Alam, Md. Golam Rabiul; Reza, Md. Tanzim
    General aviation records a fatal accident rate of approximately one per 100,000 flight hours, with loss-of-control and stall events remaining leading preventable causes. Existing flight safety systems rely on fixed expert-defined thresholds and cannot detect complex multi-sensor anomaly patterns or identify specific event types in real time. This thesis proposes a lightweight two-stage deep learning framework for real-time aviation anomaly detection and multi-label event classification on raw flight sensor data from the NGAFID General Aviation Training Set. Stage 1 uses an ensemble of two novel Transformer architectures, the Cross-Sensor Patch Transformer (CSPT) and the Hierarchical Cross-Sensor Transformer (HiCST), each incorporating a cross-sensor multi-head attention module that explicitly models inter-sensor dependencies before temporal processing. Root Mean Square ensemble fusion achieves an anomaly-class F1-score of 0.8815, recall of 0.9076, and AUPRC of 0.9523 on a test set with a 339:1 class imbalance ratio. Stage 2 uses MHANet, which introduces per-sensor independent linear projections to classify each anomalous timestep into any combination of ten simultaneous event types, achieving a macro F1 of 0.9563 and subset accuracy of 0.9627. The complete pipeline runs in 6.08 milliseconds per sensor reading on a standard CPU, confirming real-time feasibility. Both stages outperform all established deep learning and classical machine learning baselines while using significantly fewer parameters, demonstrating that domain-aware architectural specialization consistently outperforms general-purpose approaches for aviation safety monitoring.
  • Thumbnail Image
    Item
    WasteRefine: boundary-aware semantic segmentation of waste materials using a DINOv2 backbone with multi-scale feature fusion decoder
    (BRAC University, 2026-04) Khan, Talha Islam; Das, Trisha; Iqbal, Md. Ahnaf; Tawseef, Farhan; Alam, Md. Golam Rabiul; Datta, Nirjhor
    The rapid increase in world waste production needs smart, data-driven frameworks for efficient material identification and sustainable resource management. Intelligent recycling systems and waste materials spontaneous segmentation often lack behind due to scarcity of proper annotated datasets, visual ambiguities and severe class imbalancement of rare objects. The research aims to propose WasteRefine, utilizing DINOv2 Vision Transformer backbone with boundary aware semantic segmentation and multi scale feature fusion decoder for waste materials. To capture and accumulate the global context, an advanced dense predictive transformer is used consisting top-down fusion of features, Pyramid Pooling Module, Squeeze and Excitation channel attention and boundary composition component, for the proper identification of cluttered, deformed and visually ambiguous waste objects. The paper also introduces WasteRefine dataset consisting of 2,213 annotated images across four different categories: paper, soft plastic, rigid plastic and metal, marking it as the first waste semantic segmentation dataset from Bangladesh which contains visuals across various regions and annotated precisely. The proposed framework is rigorously evaluated on three different dataset WasteRefine, ZeroWaste-F and SpectralWaste (RGB) and assessed across notable published baselines. The ViT-B achieved 96.64 ± 0.16% mIoU on WasteRefine dataset, 61.94 ± 0.84% mIoU on extremely class imbalanced and deformed ZeroWaste-F dataset and 70.73 ± 0.10% FG mIoU on SpectralWaste beating all the published reports. Competitive results of the ViT-S variant with only 25.16M parameters demonstrated efficient parameter count without severe performance degradation.
  • Thumbnail Image
    Item
    An efficient technique for real-time transformation of 2D to 3D images with GPU using CUDA programming
    (BRAC University, 2026-02) Mahmud, Sadat; Mustafa, Md. Rana; Mitra, Ananda; Shanto, Sajjad Hossain; Chowdhury, Mohammad Nazibul Bashar; Alam, Md. Ashraful
    This thesis presents a complete 2D to 3D reconstruction system designed to run reliably on a low computational powered PC, where GPU memory, host memory, and disk bandwidth impose strict constraints. The pipeline begins with large-scale synthetic data generation from ShapeNet models, producing aligned RGB and depth observations for supervised learning. A ResUNet18 based monocular depth network is trained in LibTorch using a mask-aware objective to promote numerical stability and reduce invalid-depth regions in the predicted maps. To ensure continuous training without data starvation under limited resources, the system is implemented a producer consumer scheduling system design: a producer renders and stages batches to fast local storage, consumers stream and pre-process shards into the training loop, and a destroyer reclaims storage deterministically once a batch is fully consumed. This design bounds disk usage, prevents host RAM accumulation, and decouples rendering from training so the GPU remains saturated even when CPU-side work fluctuates. After inference, predicted camera-centric depth is lifted into explicit 3D geometry using CUDA-accelerated reconstruction, enabling dense point cloud and grid-mesh generation at image resolution with minimal overhead. The system is evaluated using runtime traces (GPU utilization, GPU/host memory, and CPU load) alongside standard depth estimation metrics aggregated across training batches, demonstrating sustained execution, stable memory behavior, and reconstruction-ready depth quality on resource-constrained hardware.
  • Thumbnail Image
    Item
    TRACER: task-aware risk-adaptive architecture for continual edge learning
    (BRAC University, 2026-02) Islam, Md.Hasibul; Fuad, Mir Muhammad; Tanzin, A.B.M. Fahim Hasan; Rahman, Md. Khalilur
    Current computer-vision architectures are being deployed as long-lived services, especially on edge and on-device platforms, where input distributions change with changes in environment, users, sensors, and class frequencies. In these cases, to achieve sustainable performance, continual learning is required. Also, we need to keep in mind that the process needs to be feasible under strict constraints like latency and memory. Previous experience demonstrates that device-centric measures of deployment efficiency should be used instead of proxy metrics like FLOPs, and that tail latency (e.g., p95) is a more constrained measure of deployment efficiency than mean latency. At the same time, full neural architecture search (NAS) is generally too costly to integrate into a repeated learning loop, motivating restricted, hardware-aware search strategies. This thesis presents TRACER: Task-aware Risk-adaptive Architecture for Continual Edge leaRning, a deployability-oriented continual learning pipeline that keeps a fixed feature backbone and repeatedly selects and adapts a lightweight MLP classifier head. The system follows a restricted design space with a NAS-inspired controller and a Net2Net-optimized evolutionary population so that it can adapt efficiently. The stability between tasks is ensured through risk-aware exemplar rehearsal (high-risk samples are prioritized) and knowledge distillation. Experiments on Split CIFAR-100 (10 tasks x 10 classes) and CIFAR-10 (5 tasks x 2 classes) report class-incremental (CIL) and task-incremental (Task-IL) performance. Our proof-ofconcept implementation has a final CIL mean accuracy of 0.8145 and average forgetting of 0.0341, and TIL has a final mean accuracy of 0.970 on CIFAR-10. Also, in CIFAR-100, we got 0.6136 final CIL mean accuracy, and TIL has a final mean accuracy of 0.9055 with 0.0451 forgetting. These results, which are derived from a single deterministic run, demonstrate that Lagrangian-relaxation-based constraintaware head selection, combined with risk-sensitive stabilization, provides a practical accuracy-feasibility trade-off for continual learning under explicit latency targets.
  • Thumbnail Image
    Item
    PAMM: pathway-aware masked representation learning for interpretable multi-cancer prediction
    (BRAC University, 2026-01) Chowdhury, Chandrima Roy; Rodoshi, Zarrin Tasnim; Surovi, Sumaiya Hossain; Hasan, Labib; Chakrabarty, Amitabha
    In this thesis, PAMM, a new paradigm of interpretable multi-cancer prediction based on Pathway-Aware Masked Representation Learning is introduced. To tackle the challenge of the ‘Small n, Large p’ of transcriptomics it is our holding that we apply the rigorous seven-stage pipeline of preprocessing (i.e. Log2 transform, ANOVA filter, Lasso regularization and Recursive Feature Elimination) to reduce the original high-noise 57,750 genes in Breast, Lung, GBM, and HC samples to a high-signal feature set. The basic architecture goes beyond the usual deep learning of black boxes by incorporating biologically relevant priors of KEGG 2021 Human library in a self-supervised masking scheme. In contrast to stochastic masking, the pretraining phase of PAMM uses a Pathway-Aware Masking logic where complete sets of functional genes are zeroed, requiring the model to recreate missing biological units and learn complicated inter-pathway relationships. The latent representations of the model are optimized with Optuna, and the statistical robustness is verified with twenty independent iterations, and the latent representation is further interpreted with Single-sample Gene Set Enrichment Analysis (ssGSEA). The resulting visualizations of mean pathway activity indicate that PAMM is able to capture different, clinically viable biological signatures of each cancer type. PAMM provides a clear and very precise diagnostics platform of precision oncology by filling the gap between high-dimensional self-supervised learning and functional biology. Along with closed-set multi-cancer, PAMM is also explicitly tailored to open-set recognition. Through a combined study of softmax confidence and latent space distances from class centroids, the framework can discard samples that do not adhere to any known cancer manifold. This allows the certainty of identifying unknown or non-cancerous gene expression patterns, which is very essential when it comes to a real-life clinical implementation in which unobservable conditions are the norm. This two-fold feature sets PAMM apart from the traditional classifiers and guarantees the accuracy of the diagnosis and its safety
  • Thumbnail Image
    Item
    ProtReason: a reasoning-based framework for interpretable protein function prediction
    (BRAC University, 2025-06) Ayon, Sartiz Alam; Orin, Alvi Sakib; Biswas, Arpon; Fahad Al Shahid; Shahriyer, Shaikh Faiyaz; Sadeque, Farig Yousuf
    Understanding how protein sequence determines function remains a central challenge in computational biology. While some protein language models have advanced function prediction but most of them produce outputs without any justification or explainability. Protein function can be justified by connecting biological evidence to functional conclusions. We present ProtReason: A reasoning-augmented framework that generates interpretable protein function predictions with structured reasoning traces. In this study, a curated dataset of 87K proteins is constructed which is enriched with protein domain motifs, localization predictions and structural features transformed into reasoning traces linked to functional labels. ProtReason employs a two-stage architecture that first aligns protein sequence embeddings with textual representations and then generates structured outputs including reasoning traces, functional descriptions, and confidence scores. Compared to a sequence-tofunction baseline without reasoning, ProtReason achieves significantly improved BERT F1 scores, demonstrating the benefit of incorporating reasoning prior to function prediction. A systematic ablation study with 16 model variants shows the best design principles: a single unified reasoning path is better than a multi-step chain of reasoning and generating reasoning before function prediction yields superior performance. ProtReason performs competitively on standard benchmarks while providing biologically interpretable explanations with calibrated confidence estimates.