A multi-scale feature fusion facial emotion recognition method based on dynamic attention guidance

In this paper, a novel deep neural network, DynamicEmoteNet, is proposed for solving face emotion recognition tasks in complex scenes. It combines convolution, channel attention, spatial attention mechanism and multi-scale residual network. The Dynamic Channel Spatial Attention module adaptively enhances the feature expression of the key region of face emotion and suppresses the background interference through the dynamically generated channel weights and spatial weights. The Multi-Scale Residual Block module, on the other hand, simultaneously captures global expression structure and local detail features through parallel multi-scale convolutional branching, and utilizes a dynamic attention fusion mechanism to filter cross-scale valid information and enhance the robustness of the model to expression intensity, illumination changes, and facial occlusion. We conducted systematic experiments on three public datasets: AffectNet-7, FER2013 and RAF-DB, and the method proposed in this paper shows different degrees of performance improvement on each dataset compared to mainstream methods.