Reconhecimento de expressões faciais usando deep learning: uma análise comparativa sistemática em cenários de mundo real

Abstract

This work presents a systematic investigation into the performance and robustness of deep learning architectures applied to Facial Expression Recognition (FER) in real-world scenarios. The central motivation lies in the importance of emotion recognition as a fundamental component of human-computer interaction, with applications ranging from telehealth to social robotics, focusing on performance in real-world scenarios. The main objective is to analyze the impact of the domain shift phenomenon on the transition from controlled laboratory data to dynamic real-world environments, as well as to mitigate the effects of occlusions and class imbalance. To this end, a comparative evaluation of spatial feature extraction models was conducted, including the ResNet50V2, EfficientNetB0, and ConvNeXt Tiny networks, as well as a temporal hybrid model composed of a ConvNeXt architecture integrated with an LSTM recurrent network equipped with attention mechanisms. Spatial modeling experiments revealed the superiority of the ConvNeXt Tiny model pre-trained with large datasets, which achieved an overall accuracy of 87.16% on the RAF-DB dataset, highlighting the crucial role of the statistical normalization step and sequential pre-training. In contrast, cross-validation tests revealed a drastic 53% drop in model performance when exposed to domain transfer between RAF-DB and CK+ (with accuracy declining to 34.05%), highlighting the conceptual fragility in the face of variations in poses and occlusions. In temporal experiments with the CK+ dataset, the hybrid architecture demonstrated high performance under strict subject-independent validation, reaching 95.14% overall accuracy when tracking facial muscle dynamics to the peak of expression. Additionally, ablation studies focused on class imbalance indicated that simple weighting methods in the loss function (Class Weights) outperformed synthetic interpolation strategies in latent space (SMOTE) and pixel space (MixUp), preventing overfitting and preserving subtle facial distortions. It is concluded that, although modern deep architectures exhibit high hierarchical representation capabilities, practical generalization is still severely limited by domain shift, highlighting the future relevance of domain adaptation and explainability techniques.