Comparative between RESNET-50, VGG-16, Vision Transformer y Swin Transformer for facial recognition with mask occlusion

• Autores: Tafur Acenjo, Brenda Xiomara, Tello Pariona, Martin Alexis, Escobedo Cárdenas, Edwin Jhonatan
• Formato: artículo
• Fecha de Publicación: 2023
• Institución: Universidad de Lima
• Repositorio: Revistas - Universidad de Lima
• Lenguaje: español
• OAI Identifier: oai:ojs.pkp.sfu.ca:article/6361
• Enlace del recurso: https://revistas.ulima.edu.pe/index.php/Interfases/article/view/6361
• Nivel de acceso: acceso abierto
• Materia: face recognition; RESNET-50; VGG-16; Vision Transformer; Swin Transformer; reconocimiento facial

Face recognition has become relevant in the search for non-physical contact solutions in enclosed spaces for identity verification in the context of the SARS-CoV-2 pandemic. One of the challenges of face recognition is mask occlusion which hides more than 50 % of the face. This research evaluated four models pre-trained by transfer learning: VGG-16, RESNET-50, Vision Transformer (ViT), and Swin Transformer, trained on their upper layers with a proprietary dataset. The analysis obtained an accuracy of 24 % (RESNET-50), 25 % (VGG-16), 96 % (ViT), and 91 % (Swin) with unmasked subjects. While with a mask, accuracy was 32 % (RESNET-50), 53 % (VGG-16), 87 % (ViT), and 61 % (Swin). These percentages indicate that modern architectures such as the Transformers perform better in mask recognition than the CNNs (VGG-16 and RESNET-50). The contribution of the research lies in the experimentation with two types of architectures: CNNs and Transformers, as well as the creation of the public dataset shared with the scientific community. This work strengthens the state of the art of computer vision in face recognition by mask occlusion by illustrating with experiments the variation of accuracy with different scenarios and architectures.