Lightweight Facial Expression Recognition Algorithm based on POSTER++

Facial expression recognition is to classify human facial expressions automatically from digital images. We propose a more lightweight network based on the POSTER++. By introducing three channel attention modules (ECA-Net network) to the backbone network, the feature maps with larger weight values are selected for subsequent stages. This simple strategy reduces the amount of computation while increases the accuracy slightly instead. The experiments on four datasets RAF -DB, Fer2013, AffectNet(cls7) and AffectNet(cls8) show that the improved network reduces 23% in computation (Flops), and 39.68% in parameters (Params), and the recognition accuracy is 82.784%, 82.171%, 58.4%, and 63.889% respectively, which is about 0.5 percentage points higher. These results indicate that the channel attention mechanism and the selection of feature maps have a positive impact on the expression recognition.