Aplicación de técnicas de visión artificial para el reconocimiento de las emociones básicas

Abstract

Abstract: This study focuses on the development of a sophisticated facial emotion recognition model based on deep learning techniques. A model ensemble method using VGG (Visual Geometry Group) architecture is employed in combination with K-Fold cross-validation to optimize both accuracy and the system s generalization capability. The FER2013 dataset was transformed using specific data augmentation techniques, such as rotations, brightness adjustments, and scaling, to address class imbalance and enhance the model s robustness and performance in handling variations in facial images. The proposed architecture is based on an ensemble model that combines the predictions of multiple VGG models trained on different training subsets, all derived from the original dataset. These predictions are then fed into a meta-model (MLP), which learns from the probability outputs of each fold, thereby optimizing the system's overall performance and leveraging the specific strengths of each base model. Additionally, an interactive web application was developed using the Django framework, enabling image uploads, automatic processing for emotion classification, and the generation of visual and statistical reports. The application allows users to compare the model s predictions against ground truth labels, making the analysis more intuitive and user-friendly. Experimental results with various models indicate that cooperative learning can significantly improve emotion detection accuracy. Furthermore, the combination of different models proves highly effective in real-world conditions, making this approach applicable to disciplines such as psychology, human-computer interaction, and security.