In the research on facial expression recognition of deep learning models, a lightweight facial expression recognition method based on convolutional attention is proposed in this paper to solve the problems of large dataset demand and high hardware configuration requirements. First, the model parameters are decomposed and convolved for dimensionality reduction. Then, the convolutional attention mechanism module is embedded in the model to improve its feature extraction ability. For the problem of category imbalance in a dataset, the model is optimized using a cost-sensitive loss function. Finally, the model is pretrained on a face recognition dataset before performing a facial expression recognition task to improve the model’s ability of extracting facial features. Experiment results show that the method effectively reduced the model complexity while maintaining a high level of detection effect along with having strong practicability.