LKRNet: a dual-branch network based on local key regions for facial expression recognition

The task of facial expression recognition (FER) is riddled with many challenges, such as face occlusion, head posture, illumination angle, and intensity. Due to the development of deep learning and large FER datasets in recent years, most methods have achieved notable success. This paper aims to solve the problem that general classification models are difficult to distinguish, for some easily confused expressions (such as anger and surprise). To this end, we make two contributions in this paper: (1) The model extracts weighted local key regions as local information on the final feature maps, and fuses the global information for multi-task recognition. (2) Triplet loss function is used to make the intra-class feature distance significantly reduced from the inter-class feature distance. It can enhance the discriminability of features while fitting the sample distribution. The experiments confirm that two contributions are combined to gain another round of performance boost. For instance, the results on CK+ and FER2013 datasets demonstrate the superiority of the proposed method.