IUT Institutional Repository

IUT Institutional Repository is a digital service that collects, preserves, and distributes digital material. Repositories are important tools for preserving an organization's legacy; they facilitate digital preservation and scholarly communication.

Communities in IUT Institutional Repository

Select a community to browse its collections.

Now showing 1 - 5 of 12

Recent Submissions

Underwater Image Enhancement Based on Adaptive Color Balance and Contrast Optimization with Minimum Information Loss
(Department of Computer Science and Engineering (CSE), Islamic University of Technology (IUT), Board Bazar, Gazipur, Bangladesh., 2025-10-25) Tajrin, Tamanna
Underwater imaging suffers from severe visual degradation as a result of wavelength dependent light absorption and scattering, resulting in low contrast, color shifts, and loss of finedetails. Toovercomethesechallenges, thisstudypresentsamulti-modular underwater image enhancement framework. The first module, adaptive color cor rection, addresses color shift by compensating the RGB channels and adjusting each channel using condition-checked compensation values, effectively removing bluish, greenish, or yellowish casts. The second module, contrast optimization, enhances visibility by employing dark channel prior-based estimation along with a cost func tion that maximizes contrast while minimizing information loss. The third module focuses on fusion and visual improvement, where gamma correction adjusts lumi nance levels, unsharpmaskingimprovesedgesharpness, imagefusionandhistogram stretching redistributes pixel intensities across the full range to improve tonal balance andtexture clarity. The system was evaluated on multiple underwater image datasets —OccanDark, UIEB, UIEDP, and EUVP—covering diverse degradation conditions. Performance was measured using both subjective visual assessment and objective metrics, including PCQI, UCIQE, IE, and UIQM. Results confirm that the proposed framework comparatively outperforms existing techniques in color fidelity, contrast enhancement, and detail preservation. By addressing color casts, low contrast, and poor brightness with detail representation, the framework delivers high-quality un derwater images with minimal information loss, making it suitable for marine engi neering, underwater exploration, and scientific research
Developing modified ALBERT Model for Under-Represented Classes in Depression Detection
(Department of Computer Science and Engineering (CSE), Islamic University of Technology (IUT), Board Bazar, Gazipur, Bangladesh, 2025-10-25) Sanwar, Sk Nahid
In the aftermath of the COVID-19 pandemic, global reliance on communication plat forms has surged, presenting unique opportunities to analyze mental health trends at scale. Early and accurate detection of severe depression via social media content can facilitate timely interventions. However, accurately identifying varying levels of de pression severity from textual data remains a significant challenge. Human language is nuanced and severe depression cases are often under-represented in datasets. Lan guage models such as ALBERT, BERT,andDistilBERT have become prominent tools in sentiment analysis and mental health diagnostics, yet achieving consistently high accuracy remains challenging due to class imbalances and the complexity of natural language. In this study, a modified version of the ALBERT Base v2 model was intro duced incorporating an additional self-attention mechanism designed to enhance the detection of depression in text, which is often most underrepresented in datasets and contains complex structures. The performance of the original ALBERT Base v2, the modified ALBERT, BERT, and DistilBERT models were compared in terms of classi fication accuracy and parameter efficiency. Our comparative analysis demonstrates that the ALBERT Base v2 model outperforms BERT and DistilBERT overall, and our modifiedALBERTmodelshows significantimprovement in classification tasks to de tect severe depression, by focusing on long and complex dependency within words, using a self-attention mechanism. These findings suggest that the use of ALBERT models, particularly with modifications to the attention mechanisms, can improve classification performance in mental health analysis, offering an optimized classifi cation model, utilizing the parameter sharing nature of ALBERT. This work provides valuable insights into the effectiveness of different language model architectures for mental health diagnostics and highlights the impact of attention mechanisms in the mental health domain.
Enhanced Graph Convolutional Networks for ASD Diagnosis: A Multimodal Transfer Learning Based Approach Using Chebyshev Spectral Graph and Attention Mechanism
(Department of Computer Science and Engineering (CSE), Islamic University of Technology (IUT), Board Bazar, Gazipur, Bangladesh., 2025-10-25) Ashraf, Adnan Ferdous
Autism Spectrum Disorder (ASD) is a complicated neurodevelopmental disorder marked by variation in symptom presentation and neurological underpinnings, making early and objective diagnosis extremely problematic. This paper offers a Graph Convolutional Network (GCN) model, incorporating Chebyshev Spectral Graph Convolution and Graph Attention Networks (GAT), to increase the classification accuracy of ASD utilizing multi-modal neuroimaging and phenotypic data. Alongside, a transfer learning approach has been utilized in separate experiments to explore the possibility of knowledge transfer in GCN based architectures. Leveraging the ABIDE I dataset, which contains resting-state functional MRI (rs-fMRI), structural MRI (sMRI), and phenotypic variables from 870 patients, the model leverages a multi-branch architecture that processes each modality individually before merging them via concatenation. Graph structure is encoded using site-based similarity to generate a population graph, which helps in understanding relationship connections across individuals. Chebyshev polynomial filters provide localized spectral learning with lower computational complexity, whereas GAT layers increase node representations by attention-weighted aggregation of surrounding information. Also, transfer learning leads to a more robust feature generalizations across domains. Extensive trials illustrate the model’s superiority, reaching a test accuracy of 74.82% and an AUC of 0.82 on the whole dataset, and a transfer learning based approach achieved an accuracy of 72.89% and an AUC of 0.79, surpassing multiple state-of-the-art baselines, including conventional GCNs, autoencoder-based deep neural networks, and multi-modal CNNs. Ablation studies further demonstrate the incremental influence of each architectural component, emphasizing the usefulness of the methods using graph attention, hyper-parameter adjustment, and transfer learning. The results validate the effectiveness of graph-based deep learning for ASD classification, notably in combining multi-modal data and capturing complicated inter-subject interactions. This work takes a crucial step toward building reliable, scalable, and objective computer-aided diagnostic methods for ASD
An Explainable Federated Learning Approach for Heart Disease Classification
(Department of Computer Science and Engineering (CSE) Islamic University of Technology (IUT) Board Bazar, Gazipur, Bangladesh, 2025-10-25) Sathi, Tanjila Alam
Cardiovascular disease (CVD) is a life-threatening medical condition that affects the heart and blood vessels, leading to substantial financial and social burdens. Electro cardiograms (ECG) are extensively used to detect and treat heart diseases as these are rapid, secure, non-invasive, and cost-effective. The ECG signal has several segments, including the P-wave, QRS complex, QT interval, ST segment, and T-wave, together with the associated onset, offset, and peak points referred to as fiducial points. Ab normalities in these components can be indicative of various heart diseases. Medical professionals face challenges when using the conventional approach to ECG record ing, including comprehending the intricate nature of the ECG to interpret, dealing with signal interference, and handling the frequent coexistence of other health conditions alongside heart disease. However, accurate identification of fiducial points associated with the ECG wave is crucial for proper assessment of heart disease. To automate the diagnosis of heart diseases by analyzing ECG signals, various machine learning and deep learning classification models are used where data must be centralized. The majority of contemporary research encounters challenges due to data privacy, limited dataset, and lack of collaboration and proper explanation of the result. To address these challenges, this thesis presents an explainable federated learning model for the classi fication of heart disease using ECG fiducial features. This approach is the first of its kind to classify three distinct heart conditions: arrhythmia, ischemia and healthy states while ensuring both data privacy, as well as the interpretability of the outcome. We in troduced two models for classifying three distinct heart conditions: Federated Learning with Artificial Neural Networks (FL-ANN) and Federated learning with long-short term memory (FL-LSTM), which achieved impressive accuracies of 87% and 90%, respectively. In addition, explainable artificial intelligence(XAI) models can improve trust by providing visual interpretations of their results and decisions. The proposed explainable models highlighted the fiducial features of the ECG, including P-wave height (P-H), R-wave height (R-H), the QRS complex, heart rate (HR), QT interval, and corrected QT interval (QTc), which provide invaluable insights into the heart’s electrical activity, serve as vital markers for cardiac disorders, and significantly im prove diagnostic precision. Our explainable federated learning models facilitate data collaboration while safeguarding the privacy of sensitive information. By ensuring the explainability of the outcome, these models are designed to support healthcare profes sionals, helping them make better and more accurate diagnostic decisions.
Rapid Seismic Assessment Method for Unreinforced Masonry Buildings in Bangladesh
(Department of Civil and Environmental Engineering (CEE), Islamic University of Technology (IUT), Board Bazar, Gazipur, Bangladesh, 2025-10-25) Islam, S. M. Muhaiminul
The Rapid Seismic Assessment Method (RSAM) is an inexpensive and time-efficient method for assessing the seismic vulnerability of unreinforced masonry (URM) buildings to provide timely information for decision-making on the priority of buildings that need further seismic evaluation by vulnerability scoring. The simplicity of this method and the use of visual observations make it well-suited as a first-level assessment, thus helping to determine measures that will be taken and increasing the awareness of the public to the seismic hazard effects of URM buildings. To address the need for making vulnerable structures seismically resistant in seismically prone areas, this study focuses on URM buildings. This study investigates a systematic visual assessment method as a preliminary step towards the overall seismic evaluation of URM buildings. A rapid assessment approach has been developed, specifically tailored to the construction practices and materials of the buildings. This rapid assessment technique has been designed to assist in the development of a priority list for the separate determination of URM buildings that need additional detailed seismic analysis. Through the evaluation of the seismic capacity of each load-bearing masonry wall at different levels of a building, a rapid assessment index is developed. The technique developed in this study considers some of the critical factors that affect the seismic performance of masonry buildings, such as building type, number of stories, aspect ratio, slenderness ratio of walls, opening size, irregularities, pounding effect, construction age, and the physical condition of URM structures. To make the RSAM practical and valuable, available guidelines are used, and detailed seismic analysis is performed on several URM structures for comparison of the seismic index with the analytical and finite element models. The results from the studies show that this procedure can be used as an effective and practical means for the quick assessment of many masonry buildings located in seismically active areas. RSAM can be used to prioritize URM buildings that need further detailed seismic evaluation and to optimize the use of the limited available resources. This research not only demonstrates a practical solution for the initial seismic assessment of URM buildings but also highlights the necessity of implementing measures to protect vulnerable existing URM structures. It is crucial to take appropriate actions for the safety of the public from life-threatening situations in seismically and potentially active regions.