Cheng QianLobo Marques, Joao AlexandreJoao AlexandreLobo MarquesSimon James Fong2025-04-012025-04-012024-09-14https://dspace.usj.edu.mo/handle/123456789/637510.1145/3697355.3697382Facial expression recognition (FER) is essential for discerning human emotions and is applied extensively in big data analytics, healthcare, security, and user experience enhancement. This paper presents an empirical study that evaluates four existing deep learning models—VGG16, DenseNet, ResNet50, and GoogLeNet—utilizing the Facial Expression Recognition 2013 (FER2013) dataset. The dataset contains seven distinct emotional expressions: angry, disgust, fear, happy, neutral, sad, and surprise. Each model underwent rigorous assessment based on metrics including test accuracy, training duration, and weight file size to test their effectiveness in FER tasks. ResNet50 emerged as the top performer with a test accuracy of 69.46%, leveraging its residual learning architecture to effectively address challenges inherent in training deep neural networks. Conversely, GoogLeNet exhibited the lowest test accuracy among the models, suggesting potential architectural constraints in FER applications. VGG16, while competitive in accuracy, demonstrated lengthier training times and a larger weight file size (512MB), highlighting the inherent balance between model complexity and computational efficiency.text::conference output::conference proceedings::conference paper