Objective Micro-expression ME is the subtle facial expression that reveals one ?? s inner emotions. The number of samples for micro-expression recognition is small and the number of different categories is uneven leading to difficulty in recognition and low recognition accuracy. In view of this a model framework that can improve the accuracy of micro-expression recognition was proposed. Methods Peak frames containing more key expression information were extracted from the micro-expression video sequences. An improved residual network SE-ResNeXt-50 incorporating the SE module was used to extract features from the apex frames of micro-expressions. The SE module learned the key information in the features better. ResNeXt simplified the structure by replacing the dense structure with a sparse one by means of group convolution thus improving the recognition efficiency. At the same time the Focal Loss function was used to better solve the model performance problems caused by the imbalance of micro-expression data. Results Simulation experiments were conducted on the micro-expression dataset CASME II and it was found that the improved residual network and apex frames improved the accuracy and F1 value of micro-expression recognition. Conclusion The improved residual network and apex frames can reduce the impact caused by fewer data sets so that the model has a good fitting effect. At the same time it can mitigate the impact caused by the performance differences in different categories improve the accuracy of micro-expression recognition and have better recognition performance for micro-expression recognition.