FACIAL RECOGNITION PERFORMANCE EVALUATION WITH YOLOV8, ARCFACE, AND SVM IN A CONTACTLESS EMPLOYEE ATTENDANCE SYSTEM

Abstract

Manual attendance systems, which continue to be implemented in many institutions, are vulnerable to manipulation and require significant time. This research proposes an automated facial recognition attendance system optimized to address the unique challenges posed by CCTV cameras installed at a height of 3 meters. The system integrates three main components: YOLOv8m for face detection, ArcFace for 512-dimensional feature extraction, and a Support Vector Machine (SVM) with a Polynomial kernel for identity classification. The dataset (5 classes) was augmented using 20 augmentations per image and was split into a 70% training and 30% testing ratio. An image preprocessing pipeline, including CLAHE, denoising, and sharpening, was applied to enhance the input image quality. Experimental results demonstrate high classification performance, achieving 93.7% accuracy, 0.938 precision, 0.937 recall, and an F1-Score of 0.935. Confusion matrix and PCA analysis identified that the primary misclassification occurred between the E005_employee5 and E002_employee2 classes, correlating with feature overlap. Computationally, the system achieved a throughput of 7.2 FPS on the testing hardware. The system is proven to be accurate and functional for the attendance task, although its real-time performance (FPS) is highly dependent on hardware acceleration.