Smart Attendance System Using Face Recognition : Robustness, and Accuracy Evaluation Under Occlusions, Lighting Variations, and Facial Angle Changes

Sammanfattning

Face recognition has increasingly been adopted for automated attendance systems due to its contactless nature and potential to reduce administrative overhead. However, system performance in real-world deployments is strongly influenced by environmental conditions, facial variation, and decisionthreshold selection, which are often insufficiently analyzed in practice. This thesis presents the design, implementation, and evaluation of a prototype smart attendance system based on face recognition, developed as an end-to-end solution consisting of a web-based frontend, a FastAPI-based backend, and a lightweight database. The system performs open-set face recognition by comparing facial embeddings extracted using a pre-trained model and applying a fixed similarity threshold to accept registered users while rejecting unregistered individuals. The system is evaluated through a series of controlled experiments involving seven participants and explicit impostor attempts under varying conditions, including changes in lighting, facial pose, and facial occlusion. Performance is analyzed using confusion matrices and similarity-score distribution histograms to examine recognition robustness and error behavior across scenarios. The results show that the system performs reliably under conditions similar to enrollment, while challenging conditions primarily lead to increased false rejections rather than false acceptances. In particular, facial occlusions cause a systematic downward shift in genuine-user similarity scores, resulting in consistent rejection due to conservative thresholding rather than identity confusion. The main contribution of this work lies in its transparent, scenario-based evaluation of a face-recognition-based attendance system, highlighting how deployment conditions and threshold selection affect recognition behavior. The findings underscore the importance of context-aware evaluation and careful threshold selection when deploying face recognition for attendance applications.