Learning to Leverage Multi-Scale and Multi-Level Features for Facial Expression Recognition

Recent years, intense research are done with remarkable results on the topic of Facial Expression Recognition (FER). However, extracting meaningful contextual features and aggregating discriminative representations to improve the FER performance are still challenging. To address the above problem, we formulate a simple yet effective multi-scale and multi-level features learning neural network for FER. In our work, the proposed method consists of multiple scale-specific branches and a scale-fusion branch that aims to mine the implicit complementary advantages across different scales and levels. Experiments are conducted on the dataset FER-2013 to verify the effectiveness of the proposed model. In the end, the model produces a 70.0% accurate rate in the experiment which has a better performance than the state of arts.