Abstract
This study presents a two-level evaluation of a smart-classroom attendance infrastructure that combines an embedded edge prototype with a separate open-set face-recognition benchmark. A Raspberry Pi 4 prototype was developed to integrate facial recognition, attendance registration, LCD feedback, and servo-based access simulation. Device-level testing comprised 20 access attempts under each of three controlled conditions. The prototype achieved end-to-end correct identification rates of 95.0% under standard indoor lighting and 85.0% under reduced lighting, with mean response times of 0.8 and 1.2 s, respectively. All 20 unknown-face attempts were correctly rejected, with a mean response time of 0.6 s. The dlib and SCRFD-ArcFace 1-NN pipelines were compared offline on a separate workstation using a custom open-set recognition protocol on the LFW-deepfunneled dataset. At the fixed, non-matched operating thresholds used in this study, SCRFD-ArcFace achieved 98.3% overall accuracy, a 96.7% true positive identification rate (TPIR), and a 0.4% false positive identification rate (FPIR), compared with 83.8%, 93.1%, and 23.3% for dlib, respectively. The main contribution is a clear distinction between the complete Raspberry Pi prototype tests and the algorithmic benchmarks conducted on a workstation. The prototype provides a basis for further development and validation of secure and automated educational environments, including systems intended to support learners with special educational needs.