https://ar.iub.edu.bd/handle/11348/1050 2026-10-09T04:07:15Z 2026-10-09T04:07:15Z Sakib, Abu Musa Mukta, Sumyia Afnan https://ar.iub.edu.bd/handle/11348/1598 2026-09-21T00:15:27Z 2026-04-01T00:00:00Z

DrishtiXAI: a Bengali image captioning system with explainable AI Sakib, Abu Musa; Mukta, Sumyia Afnan This study presents DrishtiXAI, an explainable Bengali image captioning system designed to support visually impaired users through culturally relevant image descriptions and text-to-speech technology. A dataset of 18,539 Bengali image-caption pairs was developed, and two deep learning approaches were evaluated. The InceptionV3-GRU model with Bahdanau attention outperformed the EfficientNet-B4-Transformer model in reliability, achieving a mean BLEU score of 0.5157 on a separate dataset. Attention maps provided visual explanations of the image regions used to generate captions, while a mobile application enabled accessible use. A user study of 20 participants demonstrated positive ratings for clarity and supportiveness, highlighting the potential of transparent and culturally appropriate AI for Bengali-speaking communities. This thesis is submitted in partial fulfillment of the requirements for the degree of Master of Science (M. SC..) in Computer Science, 2025.

2026-04-01T00:00:00Z Bhuiyan, Zarif Wasif Rahman, Md. Ataur https://ar.iub.edu.bd/handle/11348/1597 2026-09-21T00:15:25Z 2026-08-24T00:00:00Z

Automated software requirement and design understanding using deep learning-based UML diagram recognition and LLM-based design pattern detection Bhuiyan, Zarif Wasif; Rahman, Md. Ataur This thesis presents an AI-based framework for automated software requirement and design understanding by combining deep learning for UML diagram recognition and large language models (LLMs) for software design pattern identification. Using a dataset of 6,666 UML diagram images across six categories, YOLO achieved the highest recognition accuracy of approximately 99%. For design pattern recognition, GraphCodeBERT combined with a k-Nearest Neighbor classifier achieved the best F1-score of 0.969. The findings demonstrate that AI techniques can effectively analyse visual software artifacts and source code, reducing manual effort and supporting automated software analysis, reverse engineering, and intelligent software engineering applications. This thesis is submitted in partial fulfillment of the requirements for the degree of Master of Science (M. SC..) in Software Engineering, 2025.

2026-08-24T00:00:00Z Hafiz, Zinia https://ar.iub.edu.bd/handle/11348/1568 2026-09-06T12:35:18Z 2025-12-14T00:00:00Z

A lightweight framework for implementing ISO/IEC 26550 in small and medium-sized software enterprises in Bangladesh Hafiz, Zinia This study explores the challenges faced by Bangladeshi software Small and Medium-sized Enterprises (SMEs) in adopting ISO/IEC 26550 for Software Process Improvement (SPI). Data collected from 12 SMEs revealed uneven implementation across key process areas, with high familiarity in domain requirements (up to 95%) but limited engagement in testing and validation practices. To address these gaps, the study proposes a lightweight PDCA (Plan–Do–Check–Act) framework tailored to the unique resource constraints of SMEs. Unlike traditional models such as CMMI or ISO/IEC 12207, the proposed approach emphasizes simplicity, cost-effectiveness, and role-based adaptability, making it more practical for SMEs in emerging economies. The framework supports scalable and gradual process improvement, enhances quality assurance, and facilitates stronger alignment with ISO/IEC 26550 in resource-limited environments. This thesis is submitted in partial fulfilment of the requirements for the degree of Master of Science (M. SC.) in Software Engineering, 2025.

2025-12-14T00:00:00Z Setu, Jahanggir Hossain https://ar.iub.edu.bd/handle/11348/1547 2026-08-24T08:35:25Z 2026-08-01T00:00:00Z

Declared failure over silent failure: abstention, explain ability and an abstention-aware evaluation framework Setu, Jahanggir Hossain Deep image classifiers are constructed in a way so that every input produces a label, whatever the model has learned about that input. The resulting failures are silent, e.g., a classifier confronted with an Out-Of-Distribution (OOD) or adversarially perturbed image returns a class assignment with no signal that the assignment should not be trusted. Selective classification yields a partial remedy by permitting the model to decline but its standard evaluation practice, e.g., computing performance over the accepted subset and reporting coverage separately, charges nothing for that decline. Therefore, a reported score improves monotonically as the model abstains more. This study makes two contributions. The first is empirical where a SelectiveNet architecture is compared against a matched baseline sharing the same convolutional backbone across four datasets (MNIST, CIFAR-10, chest X-ray for pneumonia detection and Cats vs. Dogs), evaluated, under Fast Gradient Sign Method (FGSM) perturbation and with Gradient-weighted Class Activation Mapping (Grad-CAM) applied to the selection head, as well as, the prediction head so that the abstention decision is itself explained. The second is the Abstention-Aware Evaluation Framework (AAEF) which extends the confusion matrix with an abstention column preserving every sample and defines three metrics on it. Boundedness, exact reduction to the classical metrics at zero abstention and correct extreme-case behavior are established and verified. Absolute error counts fell on all four datasets with missed pneumonia cases dropping from 17 to 7, though 4 to 14 deferrals were required per error avoided. Adversarial results were consistent, as well. As perturbation strength rises the baseline CNN degrades steadily, while SelectiveNet’s selective accuracy recovers because it lowers coverage and rejects the samples it cannot resolve. The optimism introduced by discarding abstentions ranged from 0.069 to 0.26 in balanced accuracy. Charging abstention as an outcome rather than an exclusion changes the reported performance of a selective classifier which is large enough to reverse a comparison that conventional evaluation would decide the other way. This thesis is submitted in partial fulfilment of the requirements for the degree of Master of Science in Computer Science, 2026

2026-08-01T00:00:00Z