Fig. 1. Composition of the Fer2013 dataset.

Fig. 2. Seven facial expression categories and their corresponding spectra. Here, the numbers 0–6 represent angry, disgust, fear, happy, sad, surprise, and neutral, respectively.

Fig. 3. Examples of (a) low-pass and (b) high-pass filtering.

Fig. 4. Architecture of the network model.

Fig. 5. (a) Dependence of recognition accuracy on low-pass filtering. (b)–(e) Confusion matrices for N =5, 12, 15, and 18, respectively.

Fig. 6. (a) Dependence of recognition accuracy on high-pass filtering. (b)–(e) Confusion matrices for N = 4, 8, 12, and 16, respectively.

Fig. 7. (a) and (b) show the dependence of recognition accuracy on low-pass and high-pass filtering, respectively, using ResNet. (c) and (d) present the dependence of recognition accuracy on low-pass and high-pass filtering, respectively, using ViT.