Using attention LSGB network for facial expression recognition

Both the multiple sources of the available in-the-wild datasets and noisy information of images lead to huge challenges for discriminating subtle distinctions between combinations of regional expressions in facial expression recognition (FER). Although deep learning-based approaches have made substantial progresses in FER in recent years, small-scale datasets result in over-fitting during training. To this end, we propose a novel LSGB method which focuses on discriminative attention regions accurately and pretrain the model on ImageNet with the aim of alleviating the problem of over-fitting. Specifically, a more efficient manner combined with a key map, multiple partial maps and a position map is presented in local relation (LR) module to construct higher-level entities through compositional relationship of local pixel pairs. A compact global weighted representation is aggregated by region features, of which the weight is obtained by putting original and regional images to the sequential layer of self-attention module. Finally, extensive experiments are conducted to verify the effectiveness of our proposal. The experimental results on three popular benchmarks demonstrate the superiority of our network with 88.8% on FERplus, 58.68% on AffectNet and 94.9% on JAFFE.