Facial expression recognition method based on dual-branch fusion network with noisy labels

As an important problem in the field of computer vision, facial expression recognition has been widely used in many fields such as human-computer interaction and emotion recognition. This paper proposes a dual-branch parallel facial expression recognition network that integrates Convolutional Neural Network (CNN) and Transformer. By consolidating local and global information obtained from the dual branches, the network effectively combines subtle local variations and overall contextual features of facial expressions. The fused feature representation demonstrates enhanced representativeness and discriminative power, leading to a notable improvement in facial expression recognition accuracy. In addressing the challenges posed by inter-class similarity and annotation ambiguity, the Erasing Attention Consistency (EAC) mechanism is employed. This mechanism effectively prevents the model from memorizing noisy samples, thereby enhancing robustness against noisy labels. Additionally, the focal loss function is utilized to rectify imbalances within data samples. Experiments conducted on RAF-DB datasets with labels containing noise reveal that the proposed methodology achieves accuracies of 87.65% and 86.05% at 10% and 20% noise rate, respectively. This substantiates the effectiveness of the proposed approach.