Yenilmez, FulyaYildiz, CananUgur, Mukden2025-02-202025-02-202023979-835035901-5https://doi.org/10.1109/ICCIT60459.2023.10441242https://hdl.handle.net/20.500.12846/175326th International Conference on Computer and Information Technology, ICCIT 2023 -- 13 December 2023 through 15 December 2023 -- Cox's Bazar -- 197664In recent years, deep learning architectures have demonstrated remarkable achievements on diverse computer vision tasks, including the recognition of facial expressions. The development of precise and robust facial expression recognition (FER) models has the potential to improve a wide range of applications in various fields, such as human-computer interaction, emotion analysis, marketing, robotics, psychology, and health care.Within this domain, several Convolutional Neural Network (CNN) architectures have been employed to yield promising results. In this paper, we undertake a comparative analysis with the latest iterations of five widely adopted CNN models for the FER task: VGG19, InceptionV3, ResNet152V2, MobileNetV2, and EfficientNetV2B1. We also explore the effect of transfer learning, specifically the difference between pretraining on VGGFace versus ImageNet. Our primary goal is to comprehensively assess these architectures under identical conditions and provide clarity on their performance in comparison to each other and to previous works using various training strategies. Our assessments allow the conclusion that increasing the size and depth of these common "backbone"models leads to only minor improvements. Augmenting these models with additional architectural elements such as attention mechanisms seems much more promising in comparison. Our results further show that pretraining on a domain-specific dataset can lead to significant improvements, demonstrating the importance of a large FER related dataset. © 2023 IEEE.eninfo:eu-repo/semantics/closedAccessCNN architecturesface expression recognitionfine-tuningneural networkstransfer learningConference Object10.1109/ICCIT60459.2023.104412422-s2.0-85187362886