Relation-aware Network for Facial Expression Recognition

Abstract

Facial expression recognition (FER) is a challenging computer vision task due to problems including intra-class variation, occlusion, head-pose variation, etc. The convolutional neural networks (CNNs) have been widely adopted to implement facial expression classification. While convolutional operation captures local information effectively, CNN-models ignore relations between pixels and channels. In this work, we present a Relation-aware Network (RANet) for facial expression classification. RANet is composed of two relational attention modules to construct relationships of spatial positions and channels. Global relationships help RANet focusing on discriminative facial regions to alleviate the above problems. The separable convolution has been applied to compute spatial attention efficiently. Experimental results demonstrate that our proposed method achieves 89.57% and 65.09% accuracy rate on the RAF-DB dataset and the AffectNet-7 dataset, respectively.