http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/1196 Sun, 12 Apr 2026 20:59:51 GMT 2026-04-12T20:59:51Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16791 Bangladeshi License Plate Detection and Recognition Using YOLO Variants and Enhanced OCR with Model Interpretability Ramit, Shahriar Sultan This study presents a sophisticated framework for automatic license plate recognition (ALPR) designed for Bangladeshi vehicle registration plates, tackling the intricacies of Bangla script and diverse environmental conditions. We employ three YOLO variants YOLOv5, YOLOv8, and YOLOv11for accurate license plate detection, yielding mean Average Precision (mAP50) scores of 0.955, 0.961, and 0.950, respectively, on a primary dataset of Bangladeshi images. Detected plates undergo meticulous preprocessing with OpenCV, encompassing grayscale conversion, adaptive thresholding, contour detection, and Gaussian blur to mitigate noise and enhance text clarity. These steps are critical to address challenges such as variable lighting, shadows, and plate degradation. A tailored Optical Character Recognition (OCR) pipeline, specifically adapted for Bangla script, achieves a character-level accuracy of 89%. The OCR modifications include enhanced character segmentation and a Bangla-specific language model to overcome the complexities of Bangla’s nonlinear script, which poses significant challenges for standard OCR systems due to its conjunct characters and intricate glyphs. The framework exhibits robustness against occlusions, non-standard plate formats, and urban environmental variability, offering a viable solution for intelligent transportation systems in Bangladesh. Comparative evaluation of YOLO variants highlights YOLOv8’s superior mAP50 and YOLOv11’s high precision, informing their suitability for real-time applications. This work establishes a foundation for scalable ALPR, with potential to enhance traffic management and law enforcement in Bangladesh. Project Report Wed, 17 Sep 2025 00:00:00 GMT http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16791 2025-09-17T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16790 Revolutionizing ACL Tear Detection Using Deep Learning Models Heem, Tarian Ahamed Anterior cruciate ligament (ACL) injuries are considered to be one of the most widespread and acute knee disorders that can require appropriate and timely diagnosis because this can be one of the conditions that will be used to make decisions regarding the treatment. In manual inspection of magnetic resonance imaging (MRI), time-consuming mistakes are involved in interobserver errors. To address this problem, we propose a hybrid deep learning model, which will be based on EfficientNet-B0 and MobileNetV2, integrated with slice-level attention pooling and late fusion in the coronal and sagittal directions. The MRNet data were preprocessed through normalization, resizing, padding of the slice and augmentation of the data so as to enhance generalization. The sampling plan was identified as a dynamic weighted sampling technique to overcome the problem of class imbalance, whereas the training was conducted with the assistance of AdamW optimization, label smoothing, dropout, mix up augmentation, and exponential moving average (EMA) tracking. The experimental results show that the hybrid CNN was more efficient than the single architectures and the overall accuracy of the hybrid model is 94.1, the precision was 0.98, the recall was 0.89 on positive cases and the overall performance was balanced. The findings of the DenseNet169, EfficientNet-B0 and MobileNetV2 are 93, 91.6 and 90 percent which proves the outstanding of the hybrid approach. The model was also tested using precision, recall, F1-score and confusion matrices, which proved its strength. In this paper, I will give an insight into how a hybrid deep learning and attention can work to identify ACL tear on MRI and offer a reliable and scalable solution to assist radiologists and improve clinical decision-making. Project Report Wed, 17 Sep 2025 00:00:00 GMT http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16790 2025-09-17T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16789 Adversarial Malware Detection: Defending AI Models Against Evasive Attacks Hasan, Md Zahid Adversarial manipulation of static malware features can derail learned detectors. This research builds and audits a complete pipeline that measures and strengthens robustness for Windows executable screening under explicit, realistic threat models. Windows binaries were represented as fixed-length feature vectors and classified by three complementary learners; a Feed Forward Neural network (FNN) with bounded inputs, a one-dimensional Convolutional Neural Network (1D-CNN) with standardized inputs, and a Gradient-Boosting (LitghtGBM) tree on raw features. Evasion was evaluated with the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) for the neural models and with a decision guided routine for the tree model. Defenses combined adversarial training, feature squeezing detectors calibrated at a fixed 1% false-positive rate, and simple probability averaging ensembling. All iteration used fixed seeds and produced saved artifacts for audit. Clean baselines on the full test set were strong: gradient-boosted trees reached 97.63% accuracy and 99.62% receiver operating characteristic area under the curve (ROCAUC), the convolutional network reached 96.37% and 99.24%, and the feed forward network reached 95.68% and 98.48%. Under iterative adversarial attacks, the convolutional network falls to 0.23% at a large projected gradient descent budget and the feed forward network to 28.43% at a moderate budget. Adversarial training improved robustness only marginally while reducing clean accuracy; detectors recovered up to 44.00% true-positive rate at a 1% false-positive rate; ensembling stabilized predictions but did not neutralize strong attacks. Project Report Wed, 17 Sep 2025 00:00:00 GMT http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16789 2025-09-17T00:00:00Z http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16788 Artifact-Aware Explainable Deep Learning for Reliable Bone Fracture Detection Foragi, Dedarul Islam Bone fracture detection in radiograph is a fundamental task in clinical practice and a correct and timely evaluation of bone fractures has a very high correlation with patient outcome. If their plates, screws and implants consist of metal, then they are likely to be misread by the clinicians and by the automated devices, and therefore their diagnostic value will be reduced. This project had a motivator, to get over this limitation, and bring us towards more accurate fracture detection and interpretability. In order to reduce reliance on artifact-induced error, we developed a methodology which integrates modern deep learning with artifact reduction and prediction based on actual fracture appearance rather than artifact-induced signals. Comparisons with models presented in the literature showed that the system was generally more accurate, reliable and clinically usable than the other models. Alongside performance gains, the study also highlights the importance of interpretability - scientific descriptions that increase clinicians' confidence by ensuring safe application to the healthcare workflow. When broadly deployed, these findings can help limit misdiagnostic error and improve decision making, provide a path forward for the practical use of artificial intelligence (AI) systems in clinical medical imaging. Project Report Tue, 16 Sep 2025 00:00:00 GMT http://dspace.daffodilvarsity.edu.bd:8080/handle/123456789/16788 2025-09-16T00:00:00Z