ELG: Emotion Recognition Convolutional Model Integrating Local and Global Facial Features

Facial emotion recognition (FER) is crucial in advancing health, security, and human-computer interaction. Nowadays, The mainstream method for FER relies on computer vision, specifically utilizing deep learning models to extract facial features, enabling the identification of emotions. However, due to the local similarities and global differences inherent in facial features, the extracted features exhibit significant variations. Existing methods primarily focus on how to extract both types of features fully, neglecting the fact that the variations between them can lead to feature dependence in classifiers, ultimately resulting in lower recognition accuracy. Therefore, this paper proposes an emotion recognition convolutional model integrating local and global facial features (ELG). It includes the multi-scale local feature extraction module (MLF) and the global feature shrinkage attention module (GFS). The MLF employs multi-scale parallel convolutions to extract local features while preserving the integrity of global features. The GFS, based on attention mechanisms and feature shrinking techniques, reduces the disparities between the two types of features, mitigating the model’s feature dependence, and thereby enhancing the model's accuracy in emotion recognition. On public datasets FER-2013 and CK+, ELG outperforms state-of-the-art emotion recognition models, with accuracy and F1-score higher by 1.2% and 2.4%, respectively. Specifically, ELG achieves an accuracy of 80.2% and an F1 score of 80% on FER 2013, while reaching an accuracy of 99.5% and an F1-score of 98.4% on CK+.