Reconocimiento de expresiones faciales mediante redes neuronales convolucionales ligeras

Abstract

Facial Expression Recognition (FER) is an active research area within Artificial Intelligence (AI) with increasing relevance in real-world applications. This work explores the development of a deep learning-based FER system focused on achieving competitive performance using lightweight architectures that are suitable for environments with limited computational resources. While high-capacity models were initially explored, their computational requirements exceeded the available hardware, prompting a shift in focus toward lightweight alternatives. The final system was built around ResNet-18 and trained using transfer learning on a hybrid dataset comprising real-world, AI-generated, and publicly available images from MMI, OULU-CASIA, EFE, FERD and AffectNet. Experimental results showed that the proposed ResNet-18 model achieved a mean accuracy of 91.74% ± 0.40% (n=3), with a maximum observed accuracy of 92.27%. EfficientNet and MobileNetV3 were also evaluated and achieved competitive accuracy levels; however, their training curves plateaued early, suggesting unstable learning and limited convergence compared to ResNet-18. The system's compact design and strong results on a diverse, resolution-consistent dataset indicate its potential for future application in low-resource settings.