Robust Face Recognition via Adaptive Thresholding and Multiple InsightFace Architectures on Customized Datasets

Advancements made in deep learning are making it easier for face recognition systems to be accurate. However, challenges remain in identifying unseen identities and detecting imposters in real-world scenarios. This study introduces a robust evaluation framework using multiple pretrained InsightFace models, including buffalo_l, buffalo_m, buffalo_s, buffalo_sc, and antelopev2. The methodology involves dataset construction, embedding extraction, cosine similarity-based identity verification, and calibration using both fixed and dynamic thresholds. A dedicated imposter dataset was created to simulate unauthorized access conditions. Experimental results demonstrate that dynamic thresholding consistently outperforms fixed settings, with accuracy ranging from 96.0 to 99.7 percent, ROC AUC scores between 0.93 and 0.96, and F1 scores up to 0.999 across multiple test conditions. Among all models, buffalo_m achieved the most stable and generalized performance. This work closes the gap between ideal and realistic evaluation by integrating imposter scenarios, threshold calibration, and multi-metric evaluation for practical deployment in biometric authentication and surveillance systems.