Reconocimiento de Emociones en Rostros Utilizando una red Neuronal Convolucional Evolutiva

Automatic facial emotion recognition has achieved significant progress through the use of convolutional neural networks; however, the increasing complexity of these models has led to higher computational costs and growing environmental impact. This study aimed to comparatively analyze the performance of different convolutional neural network architectures applied to facial emotion recognition, incorporating energy efficiency and carbon dioxide emissions as complementary evaluation criteria. Widely used pretrained models in computer vision were implemented, along with a coarse-to-fine convolutional neural network and its version optimized through genetic algorithms. Experiments were conducted using the CK+ and KDEF datasets, which include seven basic emotions, and performance was assessed using accuracy, loss, training time, energy consumption, and CO₂ emissions estimated with CodeCarbon. The results showed that the coarse-to-fine convolutional neural network achieved a consistent balance between performance and sustainability, while pretrained models exhibited differentiated behavior depending on their architectural complexity. In addition, evolutionary optimization enabled specific performance improvements, accompanied by an increase in computational cost. It was concluded that the integration of sustainability metrics allowed a more comprehensive evaluation of facial emotion recognition systems and provided evidence for the development of more efficient and responsible models.