International Journal of Science, Engineering and TechnologyAuthors: Mr. Rohit Sawant, Mr. Rahul Sawant, Dr Jasbir Kaur, Ms. Sandhya Thakkar, Mr. Suraj Kanal
Abstract: Identical twin recognition remains one of the most challenging open problems in biometric security because monozygotic twins share near-identical facial geometry, and ArcFace embedding distributions overlap within the acceptance threshold range (cosine similarity 0.70–0.85), causing conventional systems to fail. This paper proposes a Twin-Aware Adaptive Face Recognition Framework with four novel contributions: (1) an Adaptive Threshold Calibration Engine using T = μ + kσ, derived from intra-class and inter-class embedding distributions; (2) a Twin-Specific Similarity Distribution Analysis module for genetically similar subjects; (3) a structured environmental robustness evaluation with quantified per-condition mitigation strategies; and (4) a real-time ONNX Runtime deployment achieving 28 ms GPU inference latency. Experiments on 10 identical twin pairs (4,000 images, 6 conditions) demonstrate 94.6% accuracy, FAR 4.2%, FRR 5.1%, and AUC 0.96 — outperforming LBPH (78.4%), FaceNet (88.7%), and DeepFace (90.1%). The optimal k = 1.00 was validated by ROC analysis, achieving EER of 4.6%.
DOI: http://doi.org/10.5281/zenodo.20712311
