Facial mark based biometric differentiation of identical twins using dynamic feature enhancement

Abstract

Abstract This comprehensive study demonstrates an advanced machine learning framework for distinguishing identical twins using facial skin marks, achieving 96.62% cross-validation accuracy and 90.6% AUC score. The methodology incorporates four distinct hyperparameter optimization techniques (random search, Bayesian optimization, particle swarm optimization, and grid search), comprehensive statistical validation, and a robust preprocessing pipeline including PCA and SMOTE. Analysis of 74 twin pairs from 319 processed images using automated facial mark detection and multi-metric similarity assessment reveals spatial distribution patterns as the primary discriminating factor. The framework employs sophisticated feature engineering (32 $$\rightarrow$$ 15 $$\rightarrow$$ 6 dimensions) and achieves statistically significant performance ( p  < 0.001) with minimal overfitting. Random search optimization emerged as the optimal method, providing the best performance-efficiency trade-off with 90.6% AUC, 88.4% test accuracy, and the fastest execution time (31.8s). The system demonstrates production-ready computational efficiency and establishes a reliable foundation for forensic biometric applications with comprehensive statistical validation and deployment specifications. Figure 1 depicts the graphical abstract.