Deep Transfer Learning with ResNet-50 and InceptionV3 for Robust Face Recognition

Abstract

Face recognition is a vital area of research in computer vision due to its significant roles in biometric identification, security systems, healthcare automation, and crime prevention. This study aims to make a comprehensive comparison between deep learning models (ResNet 50 & InceptionV3) for face recognition through a transfer learning approach by training them across 8 datasets, including Labeled Faces in the Wild (LFW), FaceScrub, AgeDB, ORL, ORL_Augmented, GeorgiaTech, Face Recognition Dataset (FRD), and the Open Famous Faces (OF) dataset. After training, we evaluated performance using the Performance Analysis (PA) by calculating metrics (validation & test accuracy, validation & test loss, etc.). Feature visualization was done using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE). Unlike traditional methods, we evaluated the two models in real time on 15 AgeDB samples using a webcam for facial recognition. We also ran the Robustness Analysis (RA) on AgeDB to assess model sensitivity to brightness and occlusion attacks. PA findings displayed that the test accuracy for ResNet-50 ranged between (97.36-100) %, while for InceptionV3, it ranged between (95.77- 100) %, which indicates the superiority of ResNet-50 in terms of accuracy. The real-time results showed that ResNet-50 had an average confidence of 81.07%, while InceptionV3 had 74.96%. DA results showed that both models are insensitive to brightness. These findings confirm that, when trained on diverse data, the two models are highly effective for face recognition tasks in real-world scenarios.