A Comparative Study of LBPH, FaceNet, and ArcFace for Robust Face Verification Under Challenging Conditions

Face verification is an essential component of biometric authentication, surveillance, and access control systems.However, the reliability of face recognition models can decrease significantly under challenging conditions such as poor illumination, motion blur, and partial facial occlusion.Therefore, a reliable and systematic evaluation of face verification methods is important for selecting appropriate approaches for practical applications.This study employs a quantitative, experimental, and comparative research design to evaluate three face verification methods representing different stages of the field: the traditional texture-based Local Binary Patterns Histograms (LBPH), the deep metric-learning model FaceNet, and the angular-margin-based ArcFace.To ensure a fair comparison, identical preprocessing procedures, verification protocols, and evaluation settings are applied across three widely used benchmark datasets: LFW, AgeDB, and CPLFW.The models are evaluated using accuracy, ROC-AUC, Equal Error Rate (EER), False Acceptance Rate (FAR), False Rejection Rate (FRR), and computational cost.The experimental findings demonstrate that FaceNet and ArcFace substantially outperform the traditional LBPH method across the evaluated benchmarks.Among the three approaches, ArcFace consistently achieves the highest verification accuracy and demonstrates greater robustness under challenging conditions.Overall, the findings indicate that modern deep learningbased methods, particularly ArcFace, provide a more reliable and effective solution for face verification applications requiring high accuracy and robustness, although computational requirements should also be considered when selecting a method for practical deployment.