Feature selection mechanism in CNNs for facial expression recognition

Facial Expression Recognition (FER) has been a challenging problem in computer vision for many decades, mainly due to the high-level variation of face geometry and facial appearance. In this paper, we propose a feature selection network (FSN) to automatically extract and filter facial features by embedding a feature selection mechanism inside the AlexNet. The designed feature selection mechanism effectively filters irrelevant features and emphasises correlated features according to learned feature maps. Experiment results on several databases demonstrate that the FSN outperforms the AlexNet by a large margin and achieves comparable results with the state-of-the-art methods. Furthermore, the FSN also shows improved generalisation ability over the AlexNet in the cross validation experiment of different datasets.