Graph Reasoning-Based Emotion Recognition Network

Semantic information from images can be used to improve the performance of deep learning methods in recognizing human emotions. In this paper, we propose a novel framework based on graph convolutional network for emotion recognition by utilizing semantic relationships different regions. First, extract salient image regions within video frame clips using bottom-up attention module construct node features graph. Then, build graphs containing and correlations nodes network. For refinement, each feature vectors is enhanced via gated recurrent unit consisting gate memory units remove redundant information. Experimental results show that our proposed method achieves superior over state-of-the-art approaches CEAR AFEW datasets.