Facial expression recognition based on strong attention mechanism and residual network

Most facial expression recognition (FER) algorithms are based on shallow features, and the deep networks tend to lose some key features in the expression, such as eyes, nose and mouth. To address the limitations, we present in this paper a novel approach, named CBAM-Global-Efficient Channel Attention-ResNet (C-G-ECA-R). C-G-ECA-R combines a strong attention mechanism and residual network. The strong attention enhances the extraction of important features of expressions by embedding the channel and spatial attention mechanism before and after the residual module. The addition of Global-Efficient Channel Attention (G-ECA) into the residual module strengthens the extraction of key features and reduces the loss of facial information. The extensive experiments have been conducted on two publicly available datasets, Extended Cohn-Kanade and Japanese Female Facial Expression. The results demonstrate that our proposed C-G-ECA-R, especially under ResNet34, has achieved 98.98% and 97.65% accuracy, respectively for the two datasets, that are higher than the state-of-arts.