LAUNet: A Latent Action Units Network for Facial Expression Recognition

The inter-class similarity of facial expressions is one of the key challenges in Facial Expression Recognition (FER). In this manuscript, a Latent Facial Action Units Network (LAUNet) is proposed for the problem of inter-class similarity of facial expressions in FER. The proposed method recognizes subtle differences between facial expressions by learning Latent Facial Action Units Features (LAUFs). Specifically, LAUNet is composed of two parts: the Latent Facial Action Units Features Extraction Network (LEN) and the Latent Facial Action Units Selection Network (LSN). Firstly, LEN extracts LAUFs from the feature map of the backbone using the spatial attention mechanism. Then, taking advantage of the channel attention mechanism, LSN captures the latent relationships between features from LAUFs to select effective features for FER. Experiments are performed on the dataset after removing the invalid data of non-face images from the original FER2013 dataset. Compared with some previous state-of-the-art methods, LAUNet achieves the highest accuracy rate of 71.31%. Depending on the backbone, LAUNet can improve the accuracy by up to 5.46% compared to the original architecture of the backbone.