IDENTIFICATION OF MONOZYGOTIC TWINS USING SWIN-TEMPORAL TRANSFORMERS WITH ATTENTION-BASED FEATURE SELECTION FOR GRAPH-BASED CLASSIFICATION

Abstract

Monozygotic twins pose a special challenge to biometric recognition and forensic identification since they are almost identical in genetic and phenotypic appearance and the traditional methods cannot adequately differentiate them. Common methods, including STR based DNA profiling, fingerprint recognition, and iris scanning and conventional CNN or hybrid CNN-RNN networks are not capable of detecting fine spatial and sequential variation leading to lower accuracy, sensitivity and specificity in the real world. In order to address these shortcomings, Swin-Temporal-GCN model, combining Swin Transformer hierarchical spatial feature extraction, Temporal Transformer sequential dependency modeling, attention-based feature selection, and Graph Convolutional Network (GCN) effective multi-class classification, was proposed in this study. The system was applied to Python platform with a heterogeneous set of 2,235 images of faces, with 725 images of MZ twins in real-time, 200 images of non-identical twins, and 625 images of test images of different illumination, pose, and noise conditions. The findings of the experiment proved the effectiveness of the proposed method as it reached the accuracy of 97 percent, and sensitivity and specificity were respectively higher than traditional methods. This framework is useful to forensic investigators, security agencies and biometrics authentication systems and offers a scalable, high-precision and robust method of identification of MZ twins.