Deep learning-assisted facial expression recognition has been extensively investigated in sentiment analysis, human-computer interaction, and security surveillance. Generally, the recognition accuracy in previous reports requires high-quality images and powerful computational resources. In this work, we quantitatively investigate the impacts of frequency-domain filtering on spatial-domain facial expression recognition. Based on Fer2013 dataset, we filter out 82.64% of high-frequency components, resulting in a decrease of 3.85% in recognition accuracy. Our finding well demonstrates the essential role of low-frequency components in facial expression recognition, which helps reduce the reliance on high-resolution images and improve the efficiency of neural network.