Facial Recognition Using Pretrained Resnet-50 and Transfer Learning for Robust Biometric Authentication

Facial recognition is vital for security, access control, and biometric verification due to its speed, accuracy, and non-intrusiveness. Despite significant advances in facial recognition technologies, achieving robust and accurate recognition under varying facial conditions remains a challenge. This study presents a transfer learning-based ResNet-50 framework for facial recognition and evaluates its effectiveness using multiple dataset partitioning strategies to identify the optimal configuration for biometric authentication. This study develops and evaluates a ResNet-50-based deep learning facial recognition system using a refined VGGFace subset of 20,000 images from 100 individuals, implemented via Google Colab and MATLAB’s Deep Network Designer. Preprocessing included resizing, normalization, labeling, and data augmentation to boost variability and generalization. Three training configurations (70:20:10, 60:20:20, and 45:35:25) were tested, with performance measured by accuracy, precision, recall, and F1-score. The 70:20:10 split achieved the best results 90.03% accuracy, 90.03% precision, 90.27% recall, and 90.01% F1-score followed by the augmented 60:20:20 model (88.52%) and the smallest 45:35:25 model (88.02%). Consistent results validate the effectiveness of ResNet-50 and transfer learning for high-dimensional facial recognition. The framework is scalable, efficient, and suited for real-time or offline applications in surveillance, authentication, and biometric security.