https://ar.iub.edu.bd/handle/11348/624 By CSE Department 2026-09-17T11:37:10Z https://ar.iub.edu.bd/handle/11348/1585 Vispol: a real-world urban visual pollution dataset for environmental and infrastructure monitoring Ahmad, Jubaer; Akbar, Md. Faiji; Simoon, M.M. This study addresses the growing problem of visual pollution in Dhaka, Bangladesh, which includes unauthorized billboards, posters and banners, waste, exposed wires, damaged walls, and deteriorated roads. To support automated monitoring, the research introduces VisPol, a real-world dataset of 1,097 images containing 15,933 annotated instances across six visual pollution categories. Five YOLOv8 models were comparatively evaluated under identical conditions, with YOLOv8x achieving the highest F1-score and mAP@50, YOLOv8l achieving the highest precision and mAP@50–95, and YOLOv8n providing the fastest inference. The study identifies YOLOv8s as offering a practical balance between detection accuracy and computational efficiency. External validation revealed reduced performance across other datasets, highlighting challenges related to annotation differences, object scale, scene composition, and class definitions. Overall, the research provides a standardized dataset and evaluation framework for visual pollution detection, urban environmental monitoring, and smart-city management. This thesis is submitted in partial fulfilment of the requirements for the degree of Bachelor of Science (BSc) in Computer Science and Engineering (CSC), 2026. 2026-08-01T00:00:00Z https://ar.iub.edu.bd/handle/11348/1583 A two-stage u-net based hierarchical pipeline for kidney and renal tumor segmentation Dhara, Showndorjo; Anny, Aeinun Nahar This report develops a two-stage 2D U-Net framework for automatic kidney and renal tumor segmentation from CT scans using the KiTS23 dataset. The first stage segments the kidney from full CT slices, while the second identifies tumors within a kidney-centered cropped region, reducing background noise and class imbalance. Five ImageNet-pretrained encoder architectures—ResNet-18, ResNet-50, DenseNet-121, EfficientNet-B3, and MobileNetV2—were evaluated under identical conditions. EfficientNet-B3 achieved the best kidney segmentation performance, with an F1 score of 0.9541 and IoU of 0.9122, whereas ResNet-18 performed best for tumor segmentation, achieving an F1 score of 0.7195 and IoU of 0.5619. The findings indicate that encoder performance depends on the specific segmentation task and that a hierarchical 2D approach can provide strong results with lower computational requirements than 3D models. This thesis is submitted in partial fulfilment of the requirements for the degree of Bachelor of Science (BSc) in Computer Science and Engineering (CSC), 2026. 2026-08-01T00:00:00Z https://ar.iub.edu.bd/handle/11348/1582 Deep learning-based classification of leukoplakia and OSCC using histopathological images: a performance evaluation study Chowdhury, Md Sazedul Islam; Ishika, Iffat Jahan; Hasan, Maruf Al This thesis investigates the use of deep learning for histopathological classification of Oral Leukoplakia and Oral Squamous Cell Carcinoma (OSCC), two oral lesions that can be difficult to distinguish due to similar microscopic characteristics. Using the publicly available NDB-UFES dataset, six CNN architectures—ResNet-50, ResNet-101, DenseNet121, EfficientNet-B0, EfficientNet-B2, and ConvNeXt-Base—were comparatively evaluated under consistent experimental conditions. Performance was measured using Accuracy, Recall, AUROC, and MCC. ResNet-101 achieved the strongest overall performance, with 89.58% accuracy, 0.9481 AUROC, and 0.7810 MCC. Grad-CAM visualization further showed that the models focused on diagnostically relevant tissue regions. The findings demonstrate the potential of deep learning, particularly ResNet-101, as a foundation for AI-assisted oral cancer diagnosis. This thesis is submitted in partial fulfilment of the requirements for the degree of Bachelor of Science (BSc) in Computer Science and Engineering (CSC), 2026. 2026-08-01T00:00:00Z https://ar.iub.edu.bd/handle/11348/1581 Integration of Bangla voice-controlled farming robot with convolutional neural networks for smart agricultural applications Islam, MD Nadimul; Islam, Sajid-Ul; Afra, Tahsina Islam This project develops AGRIBOT, a low-cost smart farming robot that combines Bengali/English voice interaction, artificial intelligence, computer vision, and robotics to support farmers in Bangladesh. The robot uses CNN-based mango leaf disease detection and performs basic agricultural tasks such as irrigation and spraying. Built with Raspberry Pi, Arduino, sensors, motors, and pumps, the system integrates Python, TensorFlow, OpenCV, NLP, and MySQL. A human-centered design approach, including farmer surveys and interviews, was used to understand user needs. The study demonstrates the potential of voice-enabled AI agricultural robots to improve access to farming information, support timely disease detection, and promote sustainable and inclusive agriculture, while identifying challenges such as limited datasets, hardware durability, and environmental noise. This thesis is submitted in partial fulfilment of the requirements for the degree of Bachelor of Science (BSc) in Computer Science and Engineering (CSC), 2025. 2025-12-01T00:00:00Z