An Expression Recognition Model Combining Attention and Residual Networks

How to integrate multi-scale features and establish interdependencies between remote channels is a challenge for expression recognition networks. Based on this, the authors put forward a residual network on PSA (pyramid split attention)-ResNet, which replaces the 3 x 3 convolution in the ResNet50 residual module with PSA to draw multi-scale features and enhance the cross-channel information’s correlation availably. At the same time, to reduce the differences between similar expressions and expand the distance between different types of expressions, a joint loss function about Island Loss and SoftMax Loss has been introduced to optimize the parameters. The method proposed in this paper carried out simulation experiments with 2 datasets, Fer2013 and CK+, increasing precision rates to 74.26% and 98.35% respectively. This result further confirms that the method has yielded a great result on expression recognition compared with the cutting-edge algorithms.