Enhanced MobileNet and transfer learning for facial emotion recognition

Facial emotion recognition plays an important role in identifying the psychological state of persons. In this context, we proposed an efficient system for facial emotion recognition based on hybrid MobileN et and Residual block architecture. This system proceeds by eliminating irrelevant images and cropping the remaining ones on face region. Moreover, we applied both under-sampling and SMOTE algorithms to overcome the problem of unbalanced dataset. On the other hand, several techniques were applied to prevent overfitting such as early stopping, mini batch shuffling and focal loss. The experiments were done on the public dataset Fer2013 based on transfer learning technique and showed very promising results that achieved the accuracy of 95.64%.