COMPARAISON DE CNN FROM SCRATCH ET D'EFFICIENTNETB0 POUR LA RECONNAISSANCE FACIALE EN CONTEXTE DE CONTROLE D'ACCES

Abstract

Physical access security in professional environments still relies largely on traditional devices (badges, PIN codes, keys) whose reliability and resistance to fraud are limited. Facial recognition by Convolutional Neural Networks (CNNs) offers a non-invasive biometric alternative, but its performance strongly depends on the training strategy. In this study, we compare two approaches for facial identity recognition applied to access control: a CNN designed and trained from scratch, and the EfficientNetB0 model pre-trained on ImageNet and fine-tuned via a three-phase progressive transfer learning strategy. Experiments are conducted on a subset of 100 identities (7,359 images) extracted from the CASIA-WebFace dataset, with an 80/20 train/validation split. Evaluation relies on global accuracy and on operational access-control metrics: False Acceptance Rate (FAR), False Rejection Rate (FRR), and Equal Error Rate (EER). Results indicate better performance of EfficientNetB0 on the studied subset, with an accuracy of 83.4%, a FAR of 0.21%, a FRR of 6.8%, and an EER of 3.51%, compared to 61.2%, 1.14%, 18.3%, and 9.72% respectively for the CNN from scratch. These gaps illustrate the value of transfer learning for facial recognition when data and computational resources are limited. They suggest the usefulness of lightweight pre-trained architectures for the design of biometric access control systems, subject to evaluation under operational conditions.