Enhanced Efficiency and Productivity through AAMS

Abstract

The Traditional attendance management systems, which rely on human operations or RFID-based solutions, frequently struggle with scalability, accuracy, and efficiency. This thesis proposes an Automated Attendance Management System (AAMS) that employs a customized YOLOv9-C model for real-time facial recognition via deep learning. The model's performance is significantly improved by adding Squeeze-and-Excitation (SE) blocks and the Complete Intersection over Union (CIoU) loss function. On a custom dataset, the baseline YOLOv9-C model had 86.2% precision and 84.9% recall, with a mean Average Precision (mAP) of 89.9% at IoU threshold of 0.5. However, the revised YOLOv9-C(M) model demonstrated significant gains, including a mAP of 93.8%, as well as improved precision (94.1%) and recall (96.6%).

These improvements can be due to the introduction of SE blocks, which promote feature recalibration, and the CIoU loss function, which maximizes bounding box localization and increases detection accuracy even in tough conditions such as occlusion or dimly lit areas. The improved YOLOv9-C model consistently outperforms the existing YOLO models (YOLOv5, YOLOv7, and YOLOv8s), according to a comparison study. The mAP for YOLOv5 was 80.2%, YOLOv7 was 89.1%, and YOLOv8s was 91.4%. In contrast, the upgraded YOLOv9-C model outperformed the others, with greater robustness, precision, and recall.

The system employs a one kind of custom dataset to evaluate the model's performance in some scenarios and settings, as well as to ensure trustworthy workforce detection in diverse contexts. By automating the attendance process, this technology reduces errors, saves administrative time, and promotes institutional efficiency.