Attention Bilinear Pooling for Fine-Grained Facial Expression Recognition

Subtle differences in human facial expressions may convey quite distinct sentiment, which makes expression recognition a challenging task. The previous studies has achieved good results on the regular expression datasets, but it shows poor recognition accuracy and robustness for facial images with small discrimination. The high dimensional space representation of bilinear model can perceive small distinctions among images, which is crucial to the fine-grained facial expression categorization. Hence, we propose to utilize the bilinear pooling to enhance the discriminate capabilities of the deep convolution networks. Meanwhile, with the aid of attention mechanism, the roles of the important spatial positions in the feature map are highlighted. Finally, With extensive experiments, we demonstrate that our model can obtain competitive performance and robustness against state-of-the-art baselines on Fer2013 and CK+ datasets.