FACIAL EXPRESSION RECOGNITION SYSTEM BASED ON GAN NETWORK DATA AUGMENTATION

Abstract

In recent years, facial expression recognition based on deep learning networks has developed. However, the deep model needs enough labelled data for training. Currently, the amount of data in most facial expression recognition databases is very limited. In order to solve the problem of insufficient data sets and unbalanced data sets, this paper designs a GAN-based method to generate more images to expand the training set and solve the problem of unbalanced data sets. In experiments, CNN's mainstream framework was used to verify the effect of enhanced data. The classification accuracy of the trained model on the original FER2013 data set is 62.17%, the classification accuracy after traditional data enhancement is 66.89%, and the classification accuracy after GAN data enhancement is 69.47%. The accuracy of the original CK + dataset is 90.15%, the accuracy of the traditional data after augment is 93.6%, and the accuracy of the GAN data after augment is 95.14%. Finally, the trained model is used in the design of OpenCV-based expression recognition UI system.