An Optimized Algorithm for Facial Expression Recognition Based on Transfer Learning

For the issues of high training costs and limited generalization ability on small datasets in facial expression recognition models, a transfer learning method is proposed. This method utilizes a pre-trained ResNet50 model as a feature extractor and combines it with a customized sparse cross-entropy loss function with label smoothing to achieve classification on the CK+48 facial expression dataset. Experimental results show that the model achieves an accuracy of 98.48% on seven facial expression recognition tasks, validating the superiority of transfer learning in facial expression recognition tasks. Label smoothing helps the model learn more general feature representations, thus providing better generalization ability when facing new data and effectively preventing overfitting.