Multi-scale and Feature Fusion for Expression Recognition Networks

This paper proposes a method for face expression recognition that combines multi-scale region awareness with multi-level feature fusion. It aims to overcome the challenge of extracting features from face regions that vary greatly in scale. Firstly, the paper uses an inverse residual structure to perform feature extraction in a high-dimensional space, compensating for the network’s lack of feature extraction capability. Secondly, the paper introduces multi-scale region awareness to address the multi-scale nature of face regions, using multi-scale feature extraction modules to extract features from different receptive fields, as well as multi-scale down-sampling to reduce the feature loss caused by pooling down-sampling. Thirdly, channels and spatial attention are introduced to suppress invalid regions. Finally, adaptive multi-level feature fusion maximises feature information at different depths of the network while avoiding inconsistencies between features. Experimental results demonstrate the effectiveness of the proposed method, achieving accuracy of 73.66%, 98.98%, and 88.07% in FER2013, CK+, and RAF-DB, respectively.