Facial expression Recognition is an essential aspect of computer vision, significantly influencing human-computer interaction, education, security monitoring, and autonomous driving. However, the subtle differences between the facial expressions of different people are difficult to capture, posing a great challenge to the task of facial expression recognition and making it challenging to find a balance between maximizing model performance and optimizing resource efficiency. A method for the fusion of 3D attention module Simam and efficient channel module ECA features is proposed and combined with affinity loss function to propose a facial expression recognition model called FCF. The FCF model accurately extracts both global and local features, addressing the nuances of facial expressions and mitigating the effects of head posture changes and complex emotional states. Extensive testing on the FER2013 dataset, the RAF-DB dataset, and the ExpW dataset proved the validity of the model with accuracies of 73.54%, 90.37%, and 74.12%, respectively. These findings reveal that the FCF model effectively improves accuracy in facial expression recognition while maintaining a lightweight architecture, surpassing existing advanced methods.