Dual-path facial emotion recognition based on attention mechanism

Facial expressions play a role in quickly understanding the emotional state of others in interpersonal communication. Similarly, with the rapid advancement of technology, Facial Emotion Recognition (FER) also plays a significant role in human-computer interaction. Anger, disgust, fear, happy, sad, surprise are the basic emotions of humans. However, the current FER technology still needs to further improve recognition accuracy to meet the needs of people. This work proposes a dual-path FER network model based on attention mechanisms. The core of this model is composed of a path to obtain emotional feature maps and a path to obtain background noise feature maps. By subtracting the feature maps output from the two paths, the purpose of reducing background noise features in the emotional feature maps is achieved. At the same time, to improve the ability of each path to extract the required feature maps and better reduce background noise during subtraction, we introduced attention mechanisms to help extract key features. Ultimately, our model can achieve a good performance with a test set accuracy of $71.47 \%$ on the FER2013 dataset, indicating the feasibility of this model.