Search

• Form a large facial expression dataset by combining public datasets. • Propose a combined center dispersion loss function to achieve high accuracy. • Propose an incremental cosine annealing to obtain a diversely trained ensemble. We propose a combined center dispersion loss function to reduce the intra-class variations and inter-class similarities of facial expression datasets and achieve high accuracy in facial expression recognition. Because of the lack of data, we strategically combine four publicly available facial expression datasets for training. Moreover, we propose an incremental cosine annealing method for deploying multiple models trained with incremental learning rates and ensemble predictions for achieving better accuracy. This method also reduces the computational cost and yields ensemble predictions of varied models, instead of similar models, that are trained with the same learning rates. We train our methods using the VGGFace network and achieve an accuracy of 74.71% on the FER2013 test set. Image, graphical abstract [ABSTRACT FROM AUTHOR]