Deep facial expression, gender and age recognition based on combined center dispersion loss and incremental cosine annealing

Abstract

We propose a combined center-dispersion loss function aimed to reduce the intra-class variation and interclass similarity of facial attribute classification datasets and to achieve high accuracy in facial expression, gender and age recognition. Due to lack of data in Facial Expression datasets, we strategically combine four publicly available facial expression datasets for training and we train gender and age on the Adience dataset. Moreover, we propose an incremental cosine annealing method for deploying ensemble models of multiple models trained with incremented learning rates and ensemble predictions for a better accuracy. This method also reduces computational cost as wella s yields ensembled predictions of varied models instead of similar models that are trained on the same learning rates. We train our methods in VGGFace network and achieve an accuracy of 74.71% on the FER2013 test set,FERPLUS test dataset for Expression Recognition achieved an accuracy of 84.46%, 92.09% on the Adience gender test set and 62% on the Adience age test set. We further our work to develop a low cost and fast multitask network to simultaneously recognize facial expression, gender and age rather than training three separate models. Finally, to achieve high accuracy in the multitask network we propose a weighted sum of loss functions while training.