Abstract: Aiming at  the impact of factors such as  occlusion, pose variations and lighting on facial expression recognition, we proposed a facial expression recognition method based on local correlation and multi-scale spatial attention. Firstly, through the local correlation module, local features were combined with global features to enhance the connections between local features, thereby improving recognition performance of the model in complex environments. Secondly,  the multi-scale spatial attention mechanism was adopted to extract and fuse spatial structural information at different levels, enhancing the robustness of the model.  Experimental results show that the proposed method demonstrates superior facial expression recognition performance on the RAF-DB and AffectNet datasets, validating its effectiveness and generalization ability.

Key words: facial expression recognition, spatial attention, multi-scale network, local correlation