Facial expression recognition is currently one of the research hotspots in the field of artificial intelligence. In addition to normal natural expressions, people sometimes intentionally change their facial expressions to mask their real emotions. Masked expressions are more complex than ordinary expressions, and the number of masked expression datasets that can be used for model training is limited. The automatic recognition of masked expressions will be a new challenge. Traditional machine learning requires the manual design of feature extraction algorithms, resulting in low expression recognition rates. Deep learning requires a large amount of labeled data and has poor model performance on small sample datasets. This paper proposes a method based on transfer learning, which transfers the pre-trained weights to the model, reconstructs the classifier to complete new tasks, and combines data augmentation and regularization to improve the robustness and generalization ability of the model. It has achieved good results on the Masked Facial Expression Database (MFED). Using leave-one-subject-out cross-validation, the recognition accuracy is 64.78% for required expressions (6R), 42.16% for experienced emotions evoked by video clips (6E), and 21.21% for 36 mixed expressions (6Ex6R), which is improved by 35.61%, 57.14%, and 88.52% compared to traditional methods, respectively. The experiment deepens the research on the recognition of masked expressions, providing a possible method for the recognition of deception and lies.