Optimizing Facial Expression Recognition: From Model Visualization Insights to Attention-Enhanced Performance

Facial expression recognition (FER) has important application value in human-computer interaction and emotion computing. In this paper, we address the problem of performance differences of existing models in different scenarios, and systematically evaluate the performance of eight mainstream deep convolutional neural network models in comparison on three widely used datasets. We analyze the different models that exhibit significant performance differences on different datasets and different expression classes through model visualization, and explore the reasons that lead to these differences. Based on the best model, we deeply analyze the relationship between the model's attention focus area and recognition accuracy, and then select multiple attention mechanisms to tune the model, and observe the effect changes and performance enhancement brought by different attention mechanisms to the model. The experimental results show that InceptionNext_Tiny achieves a good balance between recognition accuracy and computational efficiency, and the model performance is further improved by 1.1%-1.4% by introducing attention mechanisms.